WEBVTT

00:00:00.000 --> 00:00:05.000
[All right, I'm hitting record right now.]

00:00:05.000 --> 00:00:06.000
[My brains are on the floor.]

00:00:06.000 --> 00:00:18.000
Good morning or afternoon or evening, wherever you are, everyone. Welcome to the 2025 Northern California Earthquake Hazards workshop. Now watch me do my [indiscernible] impression.

00:00:18.000 --> 00:00:32.000
Okay, thank you so much for joining us. Let's just get right into it and have a lot of fun to do your first round of welcoming everyone, we have Earthquake Science Center Director Christine Goulet.

00:00:32.000 --> 00:00:37.000
Hello, everyone. [Someone else is sharing. So, okay, hold on a second.

00:00:37.000 --> 00:00:42.000
Of course, we tested this before. But someone else is sharing. Oh, okay.

00:00:42.000 --> 00:00:44.000
You can do it now, Christine.]

00:00:44.000 --> 00:00:54.000
Thank you very much. [And I cannot do otherwise. Good.] Okay, so good morning, everyone, and welcome to ensued 

00:00:54.000 --> 00:01:12.000
or the Northern California Earthquake Hazard Workshop. Really excited to be here and you know, I repeat myself sometimes when it matters and this time, this matters. This is really cool. This is the event that I'm looking forward to every year.

00:01:12.000 --> 00:01:26.000
It feels like it's like a holiday science retreat and I've been following this workshop since it started in Menlo Park and it was smaller, and it grew, and now here we are.

00:01:26.000 --> 00:01:41.000
So it's organized by the Earthquake Science Center, and I want to share what our mission is. We collect a wide range of data on earthquakes, faults, crustal deformation, and so on. We conduct the research to increase our understanding of earthquake processes, occurrences, and their effects.

00:01:41.000 --> 00:02:01.000
We synthesize this knowledge into seismic hazard assessment products, and then we disseminate the data and products to stakeholders and the public to support seismic resilience. I could summarize that more simply by we're very curious and we seek answers to help people be safer and more resilient in earthquakes.

00:02:01.000 --> 00:02:27.000
That's my summary of our mission. So this workshop brings a lot of people together and I want to thank all the different speakers that come from different organizations we're collaborating with and all that. I'm really excited that everyone is willing to come and spend time with us and share their knowledge and engage in interesting discussions on earthquakes.

00:02:27.000 --> 00:02:53.000
And also I want to remind everybody, I know John will do that all the time, but stay engaged, ask questions. The chat is a vibrant place to continue exploring ideas. And a lot of the creative ideas come from being fully engaged in this workshop. So I hope you can set aside the email today and really engage. So I want to thank our group of organizers for this year.

00:02:53.000 --> 00:03:02.000
We have some new faces in the lot, but a lot of faces you've seen before. So Shane, Steve, Susan,

00:03:02.000 --> 00:03:14.000
Evan, Alessandra, Catherine, Grace, and Dave. With support from Scott and of course John Grindle, who is our fearless DJ host,

00:03:14.000 --> 00:03:20.000
the man, and then Sarah Minson,

00:03:20.000 --> 00:03:39.000
the fearless leader, really, and she's like the magnitude eight of crustal earthquakes for just about everything for the research, for her humor, for her leadership and on that note, Sarah, I pass the baton out to you and everybody please

00:03:39.000 --> 00:03:43.000
have a great time and engage.

00:03:43.000 --> 00:04:04.000
Ahhh...thank you, Christine, that was so nice. Okay, well, moving right along, I'm not going to slow us down by saying too much. I also want to thank all of our organizers, Steve DeLong, Shane Detweiler, Susan Garcia, Katherine Guns, Evan Hirakawa, Alessandra Jerolleman, Grace Parker, David Schwartz, and Anne Wein. Thank you so much. Like, this is a

00:04:04.000 --> 00:04:11.000
huge undertaking. It takes many people to make it happen. Thank you to all the organizers. Thank you to John for running everything. Thank you to Scott for running the website.

00:04:11.000 --> 00:04:24.000
You're all amazing and you make this amazing thing happen. Other than that, I want to thank everybody for participating and everybody who participated last year for taking the poll. We do take those suggestions seriously.

00:04:24.000 --> 00:04:30.000
However, we have a wonderful and diverse group of people with a wonderful and diverse set of opinions that do not congeal.

00:04:30.000 --> 00:04:35.000
The only thing that there was a hundred percent agreement on is people really want us to get the recordings posted faster.

00:04:35.000 --> 00:04:45.000
And I can't help you with that. We do scientific review, but what we will try and do this year is like try and give you little updates so that you know that we're still trying.

00:04:45.000 --> 00:04:54.000
And I will post the links to last year's in the chat right now so that you definitely have easy access to those.

00:04:54.000 --> 00:04:59.000
And then the other thing is the chat. Get in the chat.

00:04:59.000 --> 00:05:17.000
Have fun. Get involved while we have this meeting. Now, as everything else has different opinions. Some people love it. Some people find it distracting. And if you do find it distracting, you can turn off chat and John will tell you how to do that as part of his little Zoom tutorial that will be coming up 

00:05:17.000 --> 00:05:23.000
right now. So before I pass it to John, I just want to reiterate the most important thing that we need you all to do.

00:05:23.000 --> 00:05:32.000
Remind us to record. Okay, John, take it away explain how this thing works, please.

00:05:32.000 --> 00:05:44.000
Perfect job. Thank you for the intro like you always do, it's amazing. We're just going to go through some quick guidelines and make sure that we're all on the same page for what we're doing 

00:05:44.000 --> 00:06:03.000
for the next three days. Welcome to the 2025 Northern California Earthquake Hazard Workshop. USGS strives to provide a diverse, inclusive, and respectful environment because it is essential to the

00:06:03.000 --> 00:06:07.000
integrity of our organization and our science, 

00:06:07.000 --> 00:06:08.000
if anything fails to exemplify these ideals, 

00:06:08.000 --> 00:06:17.000
please alert an organizer for immediate assistance. Also, the views expressed by the speakers do not reflect those of the USGS or federal government. The presentations were recorded

00:06:17.000 --> 00:06:26.000
in early January. 

00:06:26.000 --> 00:06:30.000
A couple of guidelines muting 

00:06:30.000 --> 00:06:31.000
and unmuting yourself,

00:06:31.000 --> 00:06:37.000
we ask that everybody stay on mute unless you are speaking. You can find that down the bottom left-hand of your screen down in the control bar. Just click that audio button to turn your mute yourself or not.

00:06:37.000 --> 00:06:43.000
Share your videos. We'd love for you to have your faces shown, especially if you're speaking.

00:06:43.000 --> 00:06:47.000
Sarah mentioned the chat. Be nice, ask lots of questions, and to turn off chat.

00:06:47.000 --> 00:06:50.000
If you go down into your Zoom 

00:06:50.000 --> 00:07:01.000
control bar there is a little arrow. If you click on that tiny little arrow, there is a checkbox that says show chat previews.

00:07:01.000 --> 00:07:10.000
If you uncheck that, all of the chats will go away.

00:07:10.000 --> 00:07:11.000
If you don't like the chat,

00:07:11.000 --> 00:07:18.000
click that little button and they will go away. If you have any issues at all, you can chat with me and we'll figure it out together. If you do have a question during our talks 

00:07:18.000 --> 00:07:23.000
please raise your virtual hand again on your Zoom control bar. You can raise your virtual hand 

00:07:23.000 --> 00:07:25.000
and we'll share your video 

00:07:25.000 --> 00:07:40.000
and you can ask questions. Like I said, video is encouraged with your full name and institution. If you want to rename yourself, that is always helpful. We've already talked about all of our organizers. There they are once again.

00:07:40.000 --> 00:07:46.000
And finally, if you have any questions at all, you can send me a chat. My name in chat is N-C-E-H-W help.

00:07:46.000 --> 00:07:48.000
And you can also email me 

00:07:48.000 --> 00:07:54.000
or Shane if you have any other questions.

00:07:54.000 --> 00:08:02.000
Or if you need anything at all. I think that's all for administrative housework. I think, Sarah, that we're ready to actually 

00:08:02.000 --> 00:08:09.000
get everything started.

00:08:09.000 --> 00:08:20.000
Well, and here we are slightly ahead of schedule. Fantastic! Okay, let's get this show started. We have a great session to start things off with.

00:08:20.000 --> 00:08:38.000
We are going to start with a session on AI and machine learning and our wonderful co-moderators to guide us through this journey today are Theresa Sawi and Rob Skoumal. [WooHoo!]

00:08:38.000 --> 00:08:46.000
Yes. Thank you, Sarah. So good morning, everyone. So again, my name is Rob Skoumal, and along with Theresa Sawi, we'll be moderating this first session.

00:08:46.000 --> 00:08:50.000
So this morning we'll be looking at some machine learning. We'll be opening up the lid 

00:08:50.000 --> 00:08:59.000
Of the black box that is machine learning. We'll be looking at how machine learning works, some of its limitations, and how to use these techniques appropriately.

00:08:59.000 --> 00:09:04.000
So this morning we've got four wonderful talks by four equally wonderful people.

00:09:04.000 --> 00:09:09.000
As a reminder, during each talk, please post your questions and comments right in the chat as we go.

00:09:09.000 --> 00:09:12.000
And then following the talks, we'll have some more time for discussion.

00:09:12.000 --> 00:09:21.000
So our first presenter is Cody Rucker from the University of Oregon, and he'll be talking about using neural networks for rate-and-state friction inversions. So take it away, Cody.

00:09:21.000 --> 00:09:32.000
Hello, my name is Cody Rucker. I'm a student at the University of Oregon where I work with Dr. Brittany Erickson and today I'm going to be talking about physics-informed neural networks for great and state friction inversion.

00:09:32.000 --> 00:09:39.000
To give an overview of this talk, we're going to start with an introduction to machine learning and physics-informed neural networks. I want to give an intuition for them.

00:09:39.000 --> 00:09:51.000
I'll show our work on inversions using rate and state friction laws. And then we're going to talk about some of the advantages and limitations some future directions and then I want to give my own thoughts on machine learning as a modeling tool.

00:09:51.000 --> 00:09:56.000
This technique or process arose out of the success of deep neural networks.

00:09:56.000 --> 00:10:05.000
Which are purely data-driven method. And it's a means of incorporating physical laws or physical data into the modeling process.

00:10:05.000 --> 00:10:21.000
So starting off a feed-forward neural network. With "L" hidden layers is essentially take some input "x" of a vector and then you consider that to be your zero input. And then you're going to apply these kind of composed repeated matrix vector multiply adds.

00:10:21.000 --> 00:10:29.000
And however many of those compositions you do gives you the depth of your network. So if you do all of them, you're going to get a "L" depth neural network.

00:10:29.000 --> 00:10:39.000
Here we're using Pi Sub-I to denote a collection of activation functions. These are usually nonlinear activation functions that are used to capture nonlinearities in whatever we're trying to train the data on.

00:10:39.000 --> 00:10:46.000
Trying to train the network on. And then we use data to represent the data network parameters and biases.

00:10:46.000 --> 00:10:55.000
The way this method kind of typically is used in machine learning approaches is that you'll have some network N and some data G.

00:10:55.000 --> 00:11:03.000
And the standard approach is to define a loss function on some data set. So you'll look at the mean squared error of the difference between the network and the data as it's viewed.

00:11:03.000 --> 00:11:13.000
And you minimize this loss. The training step that minimizes loss is to compute the gradient for the mean squared error with respect to the weights and biases data.

00:11:13.000 --> 00:11:24.000
And then you use that gradient to update the datas. And the standard machine learning approach uses automatic differentiation for the back propagation step.

00:11:24.000 --> 00:11:32.000
And basically what we do in our method is we leverage this AD to compute derivatives of the network with respect to the input rather than the parameters.

00:11:32.000 --> 00:11:40.000
And the result of this is a new network that is trainable but such that it is the derivative of some original network.

00:11:40.000 --> 00:11:51.000
Give an example of how this works in practice. If we consider various If we consider a very simple neural network, which is Wx + Vt + b.

00:11:51.000 --> 00:11:58.000
Then its first derivative with respect to x is w, and its first derivative with respect to t is going to be V.

00:11:58.000 --> 00:12:08.000
Each of these are themselves a neural network that are trainable. So if we were to define the mean squared error loss as the difference between the first derivative with respect to x and the first derivative with respect to t.

00:12:08.000 --> 00:12:24.000
And minimize this means for error, what we're essentially doing is we're encoding the partial differential equation for linear injection. I wanted to give this example because this is the basis or the basic way in which our approach works is by using these derivatives of these networks to define PDE

00:12:24.000 --> 00:12:33.000
And then training the function using these derivatives. The pin is then extended by feet forward neural networks to consider physical governing laws.

00:12:33.000 --> 00:12:40.000
So here we look at nonlinear differential operator l with solution U and parameter lambda.

00:12:40.000 --> 00:12:44.000
So an example of this would be the acoustic wave equation.

00:12:44.000 --> 00:12:58.000
Giving here. And the kind of PLDR of this is that when you do this, when you want to use this approach, you substitute in a non-sots where you assume that the neural network itself is going to approximate the true solution U.

00:12:58.000 --> 00:13:05.000
And so then you're going to be applying this differential operator to a network that is trainable rather than you.

00:13:05.000 --> 00:13:23.000
And the general representation that we use here is to create these component loss functions that some of them are mean squared errors at known data points and others are mean squared errors for the PDE residual on a set of co-location points and that your data that you're using can be observable or known boundary data

00:13:23.000 --> 00:13:30.000
Or it can be both. This is kind of a rotund slide.

00:13:30.000 --> 00:13:37.000
So we present here a general diagram for what the architecture for our pins look like.

00:13:37.000 --> 00:13:42.000
We want to point out here that the networks here are the same network, but they're being hit with different operators.

00:13:42.000 --> 00:13:50.000
And in doing so, this produces different component loss functions. So you get a different component loss function on the boundary as you do on the interior.

00:13:50.000 --> 00:13:56.000
And the sum of these loss functions gives you the true loss, the mean squared error of the total.

00:13:56.000 --> 00:14:02.000
And by minimizing, this means whatever the total, you minimize the component losses and so forth.

00:14:02.000 --> 00:14:06.000
So we use this method to solve a 2D antiplane problem.

00:14:06.000 --> 00:14:14.000
For full elastodynamic waves. And what we did is we assumed that we were going to be dealing with nonlinear friction.

00:14:14.000 --> 00:14:22.000
Along the x equals zero vault. And here we're going to be using a ring state nonlinear friction, which uses an aging state law.

00:14:22.000 --> 00:14:29.000
In this case, we used a manufactured solution to make sure that we were able to encode the proper boundary data, but also check for the ground truth.

00:14:29.000 --> 00:14:48.000
And one of the things that we wanted to figure out is if we could include an additional network into our equations that basically assumed that we didn't know the a minus b parameters so the difference which determines our velocity weakening or strengthening regions. And we wanted to see if we could learn those given the solution everywhere else.

00:14:48.000 --> 00:14:56.000
So here I showed the full interface condition that we use, which is nonlinear rate of state with an additional state variable, which we also learned alongside the true solution.

00:14:56.000 --> 00:15:04.000
So this ends up giving you a coupled system of neural networks, much like the coupled system of differential equations you normally would solve.

00:15:04.000 --> 00:15:11.000
And what we found is that the mean squared error is a non-monotonically decreasing function of the training iteration.

00:15:11.000 --> 00:15:17.000
And that our connected network, which learns the difference A minus B, our friction parameters that we were curious about.

00:15:17.000 --> 00:15:25.000
Actually solved it in maybe one or two iterations and then all of the future iterations were kind of used to solve the rest of the solution everywhere else in the domain.

00:15:25.000 --> 00:15:32.000
Again, this uses the fact that we have a ground truth, but we won't really have that in practical applications.

00:15:32.000 --> 00:15:41.000
So we're working on extending this work to look at the kind of sliding block lab experiments that determine things like friction parameters and friction laws.

00:15:41.000 --> 00:15:53.000
To see if we can incorporate data and also measure whether or not this model is effectively I guess inverting for properly for the parameters that we're curious about.

00:15:53.000 --> 00:15:58.000
So some of the limitations are that it's not obviously better than traditional methods.

00:15:58.000 --> 00:16:01.000
For at least a forward problem, maybe in higher dimensions it would be.

00:16:01.000 --> 00:16:08.000
But we also have that our loss function isn't guaranteed to be convex and we aren't entirely sure how to best quantify errors.

00:16:08.000 --> 00:16:13.000
And how to quantify convergence, which are things that we normally have in traditional methods.

00:16:13.000 --> 00:16:26.000
Additionally, how do you make sense of the output, right? Model reproducibility and verification can be pretty tough for neural network methods whenever they are the weight initiation and the choice of hyperparameters can affect how they perform.

00:16:26.000 --> 00:16:34.000
And also uncertainty quantification needs more work, even though there has been some work on Bayesian pins with some error estimates.

00:16:34.000 --> 00:16:42.000
Some future directions that we've been looking at, you know, me and Brittany have talked about this, but it'd be nice if someone sat down and quantified the actual computational benefits or lack of.

00:16:42.000 --> 00:16:49.000
When comparing these pins to traditional numerical methods. And it'd be nice to see this for both forward and inverse problems.

00:16:49.000 --> 00:16:55.000
Additionally, what is the computational complexity, memory requirements for two algorithms to reach the same level of accuracy.

00:16:55.000 --> 00:17:05.000
This is a difficult question because the neural network approach might give different levels of accuracy depending on how it was initialized or what activation function you chose or architecture.

00:17:05.000 --> 00:17:23.000
And so we have some development to be done in that area and we want more work to be done in the area of uncertainty quantification for these methods. In terms of earthquake science, it'd be cool to start exploring how well we can learn multi-scale processes and develop some benchmark exercises.

00:17:23.000 --> 00:17:28.000
Also looking into sensitivity analysis, in particular with respect to well-instrumented lab studies.

00:17:28.000 --> 00:17:34.000
This is where I want to talk about some of my thoughts on the method. And the first of these thoughts is that this isn't sorcery.

00:17:34.000 --> 00:17:50.000
Right. And I wanted to use an example here to explain what I mean by this. It's that when you're modeling, you still have to make sure that the thing that you're trying to learn or convert for is reasonable. So I wanted to give an example here where if we look at the wave speed on the wave equation right if we suppose that there are two different wave speeds.

00:17:50.000 --> 00:18:01.000
Such that they both satisfy the same PDE, then it follows that you have to get that these two wave speeds are equal to each other. This tells us that the wave speed is actually an invertible quantity here.

00:18:01.000 --> 00:18:06.000
But if we were to break up the wave speed and look at material density and elastic moduli.

00:18:06.000 --> 00:18:12.000
And make the same assumptions, then at most we're going to get that they are equal to each other in ratio.

00:18:12.000 --> 00:18:18.000
But this means that you're only getting a unique solution up to a scalar quantity of the true solution.

00:18:18.000 --> 00:18:22.000
And the problem with this is that if you try to, you know.

00:18:22.000 --> 00:18:28.000
Invert for these two quantities mutually, you're going to get possibly an incorrect answer and not be able to determine what the true solution should be.

00:18:28.000 --> 00:18:32.000
So this gives us that the material density and elastic modulus are not mutually invertible.

00:18:32.000 --> 00:18:36.000
If you set one though, you could convert for the other.

00:18:36.000 --> 00:18:46.000
Additionally, the other thing I think is worth considering is the actual efficiency of these models. So here's a big slide.

00:18:46.000 --> 00:18:55.000
So in 2021, Google looked at the training carbon footprint for multiple natural language processing models. These are very large language models.

00:18:55.000 --> 00:19:03.000
And it showed that there's a wide range of carbon footprint here, like energy consumption.

00:19:03.000 --> 00:19:12.000
But that the GP23 training, GPT-3, resulted in the highest carbon footprint. And it was equivalent to that produced by powering 138 US households for a year.

00:19:12.000 --> 00:19:28.000
And they also said that only about 40% of the model's energy consumption is due to training. The other 60% is due to inference. So I use their numbers and their information to compute a smaller chart that reframes these numbers in terms of how much rent I could afford.

00:19:28.000 --> 00:19:44.000
If I had that money instead and showed that the G Shard 600B which was the kind of best performing one that they had here only took about 3.34 months to train of rent, months of rent to train, whereas GPT-3 took about 15 years of rent to train.

00:19:44.000 --> 00:19:48.000
And then there are inferences ended up being five months in 22.4 years.

00:19:48.000 --> 00:19:52.000
This shows a massive gap in the energy consumption between these two training methods.

00:19:52.000 --> 00:20:05.000
And why is that? Well, GPT-3 is a fully connected neural network and It uses the density of its connections between neurons to enable greater transmission of information. But because of this, it incurs a high computational cost.

00:20:05.000 --> 00:20:15.000
Whereas the G Shard 600B is a sparse network. If fewer connections allow it to have a much reduce computational complexity while still having far more parameters than GT3.

00:20:15.000 --> 00:20:24.000
And the thing that's difficult about these is, well, they do have different tasks. Gchart is a multilingual translation network, whereas GPT-3 is, well.

00:20:24.000 --> 00:20:31.000
Everything, right? It's kind of the big you know the big model at the time for what large language models are.

00:20:31.000 --> 00:20:44.000
And then the question becomes like how much of this energy is being wasted on, you know, inefficient training essentially Does GPT-3 need to be dense? Well, maybe it does for the task that it's trying to complete.

00:20:44.000 --> 00:20:49.000
But how much can we actually get away with with trying to reduce the computational complexity?

00:20:49.000 --> 00:20:54.000
So this gives me a few takeaways that I think are important to consider.

00:20:54.000 --> 00:20:59.000
Is that different architectures may achieve a lower carbon footprint than others without any loss of accuracy.

00:20:59.000 --> 00:21:08.000
But also it's unclear which architectures are most efficient for a given problem. So in earth sciences, in our modeling, it's not entirely sure what we should be using.

00:21:08.000 --> 00:21:16.000
And I think going forward, what we should be doing is on compute intensive models, we should be reporting the energy consumed and the CO2 equivalent explicitly.

00:21:16.000 --> 00:21:28.000
In addition to this, including the time and number of processors for training for these models is important so that we all understand the cost and it helps us move towards more efficient models as we try to develop you know proper solutions so

00:21:28.000 --> 00:21:35.000
I appreciate you all taking the time to listen to my talk today. If you have any questions for either myself or Brittany, you can reach us via email.

00:21:35.000 --> 00:21:43.000
But yeah, thank you for your time and listening.

00:21:43.000 --> 00:21:57.000
Alrighty. Thank you, Cody. So next up, we'll be hearing from Carrie Ann Bergen. She'll be talking about What all of us should know before using AI.

00:21:57.000 --> 00:22:08.000
Hi, everyone, and thanks to the organizers for inviting me to speak with all of you as part of the workshop. My name is Carrie Ann Bergen. I'm an assistant professor of data science and earth environmental and Planetary Sciences at Brown.

00:22:08.000 --> 00:22:13.000
My talk today is Beyond the Buzzwords, What Every Earth Scientist Should Know Before Using Machine Learning.

00:22:13.000 --> 00:22:28.000
I've replaced the buzzword AI with machine learning. So the subject of the talk is maybe a little bit overly ambitious in scope and could probably be an entire workshop series. So I'm going to try to focus on addressing kind of two ends of the spectrum on the subject of using machine learning for science.

00:22:28.000 --> 00:22:36.000
Talk about a couple of misconceptions and then address how some of these relate to two of the research projects going on in my group.

00:22:36.000 --> 00:22:54.000
So I'm thinking about how what data science what scientists should know when they're considering working with machine learning tools and their research. I tried to draw on sort of several different experiences, one kind of moving into geoscience applications of machine learning as a grad student focused on data science.

00:22:54.000 --> 00:23:11.000
Some of the new collaborations and attempts at forming collaborations as a junior faculty member with climate scientists where some were very successful, resulting in publications and other kind of fizzled out or nothing much came of them. And also thinking about teaching introductory machine learning to geoscience

00:23:11.000 --> 00:23:18.000
Undergrads and grad students where they're often working on a machine learning project or applying machine learning to their research for the first time.

00:23:18.000 --> 00:23:31.000
And I sort of, when I talk about When I talk to scientists who are using machine learning in their research, I get a lot of different reactions, but two that are not uncommon and kind of fall on opposite ends of the spectrum.

00:23:31.000 --> 00:23:40.000
Are those who say machine learning is going to magically solve my problem. And those who say machine learning can't be trusted in scientific applications.

00:23:40.000 --> 00:23:45.000
Models are black boxes and they don't understand physics, right? So these kind of represent two opposite reactions.

00:23:45.000 --> 00:23:51.000
The machine learning will magically solve my problem, I think might be due to our friend or perhaps nemesis, Chad GPT.

00:23:51.000 --> 00:24:07.000
Large language models like GPT, on which ChatGPT is built, can perform a wide range of tasks. And that's in part because language, which they're trying to model, is primarily designed and used by humans for a large range of kind of human interactions and communications to do a wide range of tasks.

00:24:07.000 --> 00:24:13.000
It's sort of natural that a language model is able to do a wide range of different tasks.

00:24:13.000 --> 00:24:20.000
So the impressive kind of qualities of ChatGPT, I think make people really optimistic about how these tools might advance their own research.

00:24:20.000 --> 00:24:27.000
On the other end of the spectrum, I think some of us, as we're starting to encounter more of these AI tools in our daily lives.

00:24:27.000 --> 00:24:32.000
Are starting to see the examples of failures. So this is an example I sort of saw on social media and actually checked out for myself.

00:24:32.000 --> 00:24:45.000
And screenshotted as the sort of search results. That have the AI piece where it doesn't know what temperature has trouble answering questions about what temperature water freezes at and gives incorrect results about scientific questions.

00:24:45.000 --> 00:24:49.000
So I think these kind of things give people pause with good reason.

00:24:49.000 --> 00:24:58.000
The challenge in viewing machine learning with too much skepticism and saying we shouldn't use them is that we miss an opportunity to really advance scientific research.

00:24:58.000 --> 00:25:04.000
There have been some really significant projects, one really high profile one in the field of biology.

00:25:04.000 --> 00:25:09.000
Which solved a decades long major problem in the field or not solve, but significantly advanced.

00:25:09.000 --> 00:25:15.000
Was the Alpha Fold project dealt with the problem of protein folding.

00:25:15.000 --> 00:25:20.000
So we don't want to miss the opportunity for the potential advances, but I think the skepticism that we have is.

00:25:20.000 --> 00:25:28.000
Some of our colleagues have is healthy. So if I were to try to answer the question in one slide of what should every earth scientist know.

00:25:28.000 --> 00:25:32.000
Before using machine learning, it's that machine learning isn't a magical solution. It's a tool.

00:25:32.000 --> 00:25:46.000
It's a tool that can potentially accelerate scientific research, but we have to approach it with curiosity and skepticism. And I use these terms because these are the ways that I try to get the students in my class to think about their use of machine learning, that they're approaching it with

00:25:46.000 --> 00:25:54.000
Curiosity, trying to try different techniques, trying to play with it, but also being skeptical of the results.

00:25:54.000 --> 00:26:07.000
Making sure they're valid, thinking through how things go wrong, how they're evaluating them. So kind of taking both both sides. I'm not going to talk in this about all the ways that machine learning could potentially accelerate scientific research.

00:26:07.000 --> 00:26:18.000
A lot of different tasks that we could apply machine learning to, but they're kind of review papers that I'm happy to point people to on this topic. So that won't be the focus. Instead, I'm going to talk a little bit more about this kind of

00:26:18.000 --> 00:26:23.000
Some of the ways that we can think about machine learning as a tool.

00:26:23.000 --> 00:26:29.000
So since machine learning, we want to think of it as a tool, not a solution, then the question is, when is it useful?

00:26:29.000 --> 00:26:35.000
I think some of the sort of optimists thinking about machine learning as being a magical solution.

00:26:35.000 --> 00:26:40.000
Have run into issues when they actually try to think about, well, is it useful for my task? And for some tasks it is useful and for others it's not.

00:26:40.000 --> 00:26:46.000
So if we're trying to think about where machine learning is helpful, it's helpful to think what does machine learning actually doing for us.

00:26:46.000 --> 00:26:55.000
So machine learning, the ultimate goal usually is to build a predictive model There's other things we can do with machine learning, like representing relationships and data.

00:26:55.000 --> 00:27:00.000
But generally our goal is something like building a predictive model in most cases.

00:27:00.000 --> 00:27:05.000
And these machine learning models that we're using have a structure. These are called architectures.

00:27:05.000 --> 00:27:12.000
With a lot of parameters and these architectures represent a class of possible models. So there's a large range of values these parameters can take.

00:27:12.000 --> 00:27:26.000
So we can have a number of different models that we could train with the same architecture. So finding the particular model Selecting that optimal model is going to require data to find the best possible solution.

00:27:26.000 --> 00:27:39.000
So we need task-specific information or data related to are the tasks that we're interested in specifically. And the more complex our task is, the more complex the model is that we're trying to fit this architecture.

00:27:39.000 --> 00:27:44.000
Generally the more data or information we need. And here I use the term information because usually our information is data.

00:27:44.000 --> 00:27:50.000
But it also could be domain knowledge in some cases. So where is this information coming from?

00:27:50.000 --> 00:27:57.000
Ideally, what we typically are looking for with machine learning and what you would in sort of an introductory class you would typically see is.

00:27:57.000 --> 00:28:05.000
We want labeled examples of data where we know both the input that we'd want to feed to the model and the output for some examples.

00:28:05.000 --> 00:28:12.000
So we know the right answer. We can feed those to the model and we want those to be examples for the task that we're specifically interested in.

00:28:12.000 --> 00:28:18.000
Ideally, we would also like to have a large number of these and we would like them to be diverse and high quality. So data quality is really critical.

00:28:18.000 --> 00:28:25.000
And this is sort of our ideal, but there are ways to use machine learning with other, bring in other kinds of data and information.

00:28:25.000 --> 00:28:29.000
For example, we can bring in domain knowledge into our system.

00:28:29.000 --> 00:28:47.000
And it's helpful if we are able to represent that domain knowledge in sort of a mathematical way so that we can directly integrate it into our machine learning analysis pipeline. We can also use unlabeled data related to the target task. So that might be

00:28:47.000 --> 00:28:53.000
We might only have a few data where we know the right answer, but maybe we have a lot of data where we don't and we can use that to help improve the model.

00:28:53.000 --> 00:28:57.000
Even if we don't know the right answer for every example that we have.

00:28:57.000 --> 00:29:01.000
We also can use data either labeled or unlabeled.

00:29:01.000 --> 00:29:15.000
For tasks that are related to our target task. So that could be something like synthetic data. It could be data that are a little bit similar, maybe collected by a different instrument, so on. Those might contain information that's useful for our task that we can use

00:29:15.000 --> 00:29:26.000
To help improve our learning and learn a better model. So generally, if people are asking, you know, what kind of tasks are well suited for machine learning in the sciences.

00:29:26.000 --> 00:29:35.000
There's no one kind of task, but generally I would say that if you have a task where you have A lot of data, including potentially data from related tasks.

00:29:35.000 --> 00:29:43.000
But maybe you don't have a lot of mechanistic understanding. So if you already know what your model is exactly, you have a good mechanistic model, it's efficient.

00:29:43.000 --> 00:29:49.000
Then machine learning is not really going to solve your problem. If you have a lot of data about something, but you don't actually know how to write the code to do it.

00:29:49.000 --> 00:29:57.000
That's something that machine learning is likely to do well. And one of the challenges where I think some of the people who view machine learning as a potential magical solution 

00:29:57.000 --> 00:30:06.000
to their problem is that they're running into sort of a fundamental challenge of they don't have a lot of data or information. And so they're just kind of have this hope that machine learning will solve the problem,

00:30:06.000 --> 00:30:10.000
but if you don't have data, unfortunately, it can be very challenging to apply machine learning.

00:30:10.000 --> 00:30:22.000
That said, there's a lot of research recently on a topic of scientific foundation models, so these are models large machine learning models trained on large, broad data sets 

00:30:22.000 --> 00:30:29.000
that are designed such that you can take that model and then adapt it for specific tasks where maybe you don't have as much data.

00:30:29.000 --> 00:30:36.000
So a couple of recent examples of this are a Geospatial foundation model trained on a range of Earth observation data 

00:30:36.000 --> 00:30:48.000
developed by NASA and IBM, and that's something where if you had a task involving Earth observation data, you may be able to then take what that model has learned, the features it's learned in this remote sensing data for instance and

00:30:48.000 --> 00:31:08.000
repurpose it for your own task. Similarly, Microsoft has created a foundation model for the atmosphere called Aurora for atmospheric modeling tasks, things like weather prediction or pollution prediction that you can use, so there's a sort of movement toward these sort of models that allow us to take information from related tasks 

00:31:08.000 --> 00:31:14.000
and bring those into our task through these foundation models, so it's kind of an exciting direction, I think 

00:31:14.000 --> 00:31:22.000
in AI for science. Alright, so as we try to think more about applying machine learning and scientific applications 

00:31:22.000 --> 00:31:27.000
I want to address a few kind of misconceptions that a lot of people seem to run into.

00:31:27.000 --> 00:31:33.000
One is that domain knowledge isn't important and actually, I think domain knowledge is really essential.

00:31:33.000 --> 00:31:44.000
For several different things. One is problem formulation, formulating a problem is often the hardest part, like taking the specific scientific question or task and turning it into a machine learning problem can be quite challenging.

00:31:44.000 --> 00:31:47.000
So my students often have trouble with with their first machine learning project.

00:31:47.000 --> 00:31:57.000
Data curation is also something that requires a lot of domain expertise, knowing what's the right data, ensuring that your data aren't biased, that they're representative 

00:31:57.000 --> 00:32:00.000
of the sort of population that you're interested in are all really important.

00:32:00.000 --> 00:32:10.000
And then model validation and evaluation interpretation, trying to make sure you're using the right metrics. The model is doing what you expect it to, that its behavior makes sense.

00:32:10.000 --> 00:32:16.000
These are all things that require a lot of domain knowledge. So it is essential in the machine learning process.

00:32:16.000 --> 00:32:29.000
Another kind of misconception is that machine learning is here to like replace all the physical modeling. And I think that's generally, maybe there's some people who have that view or are trying to do that, but for the most part, even as someone who's working on emulators 

00:32:29.000 --> 00:32:33.000
machine learning is really designed to complement physical modeling and not replace it.

00:32:33.000 --> 00:32:40.000
So there are machine learning models that are quite high profile these days that emulate numerical weather prediction systems for weather forecasting 

00:32:40.000 --> 00:32:53.000
emulate climate models, but these still rely on those models to generate their training and validation data. So they're not actually able to replace them, they just allow us to run these predictions say faster, make larger ensembles,

00:32:53.000 --> 00:33:01.000
these kinds of things. So they allow us to do different kinds of science and experiments than we could otherwise, but not replace physical models.

00:33:01.000 --> 00:33:07.000
And then finally, there's this view that if you want to use machine learning, the literature moves really fast there's always all these complicated models that are coming out.

00:33:07.000 --> 00:33:14.000
Somehow, in order to kind of keep up with the field or to use these techniques, you have to use the biggest models, the most state of the art architecture,

00:33:14.000 --> 00:33:28.000
but these are not always the best and especially for scientific data, often kind of smaller models that are chosen to fit the characteristics of your data often will work better. They might generalize better and require less resources like less data 

00:33:28.000 --> 00:33:36.000
and less compute to train. So you don't always have to have the biggest, most complicated state of the art model if you learn some kind of some of the basic machine learning techniques 

00:33:36.000 --> 00:33:44.000
those will often get you pretty far in your research, and so that's just something that I think is useful for people to be aware of.

00:33:44.000 --> 00:33:53.000
Now, getting back to our sort of skepticism. I think that when we're applying machine learning in the sciences, we should always kind of be thinking about how trustworthy are our results,

00:33:53.000 --> 00:33:58.000
and this is an area of academic research. I'm sure some of my colleagues in this session are talking about this topic as well.

00:33:58.000 --> 00:34:03.000
So one thing to keep in mind is that machine learning is not robust to new scenarios.

00:34:03.000 --> 00:34:14.000
This is especially critical in these climate applications, but is true across the board. If you apply it in different situations than it was trained on, it's not always going to work as well as you might expect.

00:34:14.000 --> 00:34:20.000
The models are also sensitive to data quality and noise, and they reflect the biases in the data.

00:34:20.000 --> 00:34:25.000
So it's really important to do robust testing and evaluation of the models.

00:34:25.000 --> 00:34:31.000
Your domain knowledge plays a key role in this, but that's a really important piece in making sure we understand how well the models are working 

00:34:31.000 --> 00:34:37.000
and in what cases. Physics informed approaches can help also with the making the models more robust, 

00:34:37.000 --> 00:34:43.000
but it's not necessarily going to necessarily replace kind of this robust testing and evaluation that we should do.

00:34:43.000 --> 00:34:49.000
Then there's the machine learning or black boxes concern. Machine learnings models are complex and it is challenging to interpret them,

00:34:49.000 --> 00:34:59.000
and so explainable AI is an active area of research with tools that are available to help users extract some understanding or meaning from these models.

00:34:59.000 --> 00:35:07.000
They tend to give you some limited information about what's happening. You're not going to get a full picture because the models are complex. You're going to get sort of a snapshot 

00:35:07.000 --> 00:35:13.000
of one aspect of what the model is doing, but that's often helpful for getting some insights or debugging potential problems with the models.

00:35:13.000 --> 00:35:21.000
So it's not true that they're complete black boxes, and again, this is an area of active research that also in my group we're working on.

00:35:21.000 --> 00:35:29.000
And then finally, models are not physically necessarily physically consistent, and again, there's a number of techniques in the literature for incorporating physics and domain knowledge 

00:35:29.000 --> 00:35:33.000
that we can look at, this might be hard constraints or soft constraints, 

00:35:33.000 --> 00:35:38.000
and I'll give an example that this is not always as hard as it sounds like it would be, 

00:35:38.000 --> 00:35:42.000
and so the second project I'm going to talk about just now, I'll mention that.

00:35:42.000 --> 00:35:46.000
So there's a couple of projects in my group that I want to kind of point to some of these things.

00:35:46.000 --> 00:35:51.000
How they relate to these projects. So my student Peter is working on machine learning based emulators.

00:35:51.000 --> 00:35:56.000
For the C-level contribution due to the melting of the Antarctic and Greenland ice sheets.

00:35:56.000 --> 00:36:09.000
So this is taking basically computationally expensive climate models and trying to take the inputs and outputs of those which often take weeks to run and train a machine learning model to approximate what it's doing so we can do faster experimentation and sensitivity testing.

00:36:09.000 --> 00:36:23.000
But because these kind of simulators and simulations and emulator results are used in things like the IPCC reports. We really need to understand things like the uncertainty of the models, so we're incorporating methods to quantify and disentangle the sources of the uncertainty,

00:36:23.000 --> 00:36:36.000
and the predictions. And we're also using XAI techniques like one called SHAP and doing sort of targeted experiments in order to get insights into what our model is doing and how it's making predictions. So we're able to determine things like

00:36:36.000 --> 00:36:41.000
climate forcings like service mass balance are the most important, that's what we would expect 

00:36:41.000 --> 00:36:46.000
compared to things like the specific modeling choices that were made by the different modeling groups 

00:36:46.000 --> 00:36:52.000
and that adding additional forcings beyond surface mass balance or temperature can help improve the performance 

00:36:52.000 --> 00:37:04.000
that our neural network outperforms the Gaussian process because it's more flexible and computationally efficient so we can actually give it more forcings and temporal information and so that's kind of where the advantage comes from.

00:37:04.000 --> 00:37:16.000
Another student working with me named Maria Luisa was working on a physics informed approach for downscaling or resolution enhancement of gridded climate data so we're looking at Arctic sea ice 

00:37:16.000 --> 00:37:32.000
and we want to increase the improve the resolution so the climate model is coarse resolution. we need to have higher resolution information to understand how the changing climate will impact the navigability of the Arctic. So we built a machine learning model that maps from the low resolution to the high resolution data.

00:37:32.000 --> 00:37:39.000
But there's concerns about generalization of this model under climate change so one of the things we can do is put in some physics 

00:37:39.000 --> 00:37:42.000
and here we put in a physics constraint, it's a hard constraint 

00:37:42.000 --> 00:37:46.000
so enforce exactly, which is in the high and low resolution snapshots,

00:37:46.000 --> 00:37:53.000
they should have the same volume of ice. So we're conserving the ice volume following an approach by Harder et al.,

00:37:53.000 --> 00:38:04.000
and this is a very simple thing to implement. Just as a layer where we're applying a normalization. So we're just basically adjusting the output so that these input and output ice volumes match 

00:38:04.000 --> 00:38:14.000
within each sort of local area. So this is something that doesn't require a lot of expertise or tweaking of the architecture. We just insert this layer kind of right before we make our final prediction.

00:38:14.000 --> 00:38:24.000
And so it's something that we can incorporate without having to have a lot of kind of digging into the details of the architecture.

00:38:24.000 --> 00:38:31.000
And so I'll end there just as my kind of summary. So machine learning is not a magical solution, it's a tool.

00:38:31.000 --> 00:38:44.000
But we should approach it with both curiosity and skepticism. So thank you for your time and I'm happy to take questions in the Q&A at the end.

00:38:44.000 --> 00:38:51.000
Thank you so much, Karianne. And next we have Amir Behzadan from University of Colorado.

00:38:51.000 --> 00:39:00.000
Who will be talking to us about "Building trust and fairness into AI-driven disaster damage outcomes." So let's welcome Amir.

00:39:00.000 --> 00:39:08.000
In this talk, I would like to highlight two important aspects of artificial intelligence implementation in disaster management.

00:39:08.000 --> 00:39:15.000
Specifically, I will discuss trust and fairness and how they can influence AI-driven disaster damage outcomes.

00:39:15.000 --> 00:39:32.000
What makes this conversation relevant to a large percentage of people and communities in the U.S. and across the globe is that AI has already started to dramatically affect our lives and the sentiments around the integration and use of AI has not always been positive.

00:39:32.000 --> 00:39:41.000
Here, for example, I'm showing results from a recent survey where respondents have clearly expressed feelings of nervousness towards AI products and services,

00:39:41.000 --> 00:39:53.000
and this perception has only gotten worse in recent years. If you're interested, you can find more details about this and other surveys in the Stanford University's AI Index Report.

00:39:53.000 --> 00:40:10.000
This more than anything else highlights the importance of rigorous fact-based investigation into the role of data and technology for making informed high-stakes decisions that uphold truth and strengthen the integrity of our communities.

00:40:10.000 --> 00:40:33.000
So in my research lab at CU Boulder, our mission is to design human-centered AI technologies to measure, map, and communicate the impacts of disasters on various communities. The term human-centered plays a critical role in this regard because it requires us to transcend the traditional metrics of technology effectiveness such as accuracy, reliability, latency, and efficiency and

00:40:33.000 --> 00:40:39.000
move to our performance indicators that take into account the user needs, values, and cultural norms.

00:40:39.000 --> 00:40:46.000
Some of these metrics are listed here and include accessibility, inclusivity, fairness, safety, transparency 

00:40:46.000 --> 00:40:54.000
and of course, ethical use. This slide shows a summary of some of the recent projects that my lab has completed 

00:40:54.000 --> 00:41:02.000
or is currently working on, all of which sit at the intersection of society built in natural environments and technology.

00:41:02.000 --> 00:41:10.000
If you're interested in learning more about any or all of these projects, please contact me and I will be happy to provide further information.

00:41:10.000 --> 00:41:16.000
But for the remainder of this talk, let me introduce the concept of trust in AI.

00:41:16.000 --> 00:41:21.000
As a construct that describes the willingness of a person or a group to rely on others, an artifact,

00:41:21.000 --> 00:41:32.000
or a technology. This definition is also in line with that of MIST, where trust is expressed by user trust profile and perceived system trustworthiness.

00:41:32.000 --> 00:41:38.000
One UTP deals with the characteristics and backgrounds of the user of a technology.

00:41:38.000 --> 00:41:44.000
PST deals with the humans assessment of the trustworthiness of the system itself.

00:41:44.000 --> 00:41:53.000
Considering high and low extremes of UTP and PST values, one can imagine four different trust scenarios that I have visualized here in this circle of trust.

00:41:53.000 --> 00:41:58.000
In one end, we have high values for both UTP and PST.

00:41:58.000 --> 00:42:01.000
And in the other end, we have no values for UTP and PST.

00:42:01.000 --> 00:42:06.000
And of course, trust is not a static phenomenon and evolves over time.

00:42:06.000 --> 00:42:17.000
Either in the positive or negative direction. So although we may start from a low end of the spectrum, trust calibration may lead to a higher degree of trust or vice versa.

00:42:17.000 --> 00:42:26.000
NIST further breaks down PST into two primary components. The first one is the perceived technical trustworthiness or PTT.

00:42:26.000 --> 00:42:39.000
And the second one is the user experience or UX. PTT encompasses the most known list of metrics using which we measure the technical competence of the technology. Metrics such as accuracy,

00:42:39.000 --> 00:43:00.000
reliability, explainability, security, and privacy all fall under this category. UX, on the other hand, tries to quantify the user experience with the technology. The idea here is that even the most technically competent system may not be trusted if you don't let humans interact with it properly and sufficiently.

00:43:00.000 --> 00:43:08.000
So to this end, UX metrics include efficiency, which is a measure of time, effectiveness, which is a measure of the number of errors made,

00:43:08.000 --> 00:43:17.000
and satisfaction, which is a measure of affect or how frustrated, engaged, or joyful the user feels when interacting with the technology.

00:43:17.000 --> 00:43:33.000
To this, I would also add relatability which describes the extent to which the technology resembles humanness and therefore, can influence the user experience in a positive or a negative way.

00:43:33.000 --> 00:43:48.000
Here is a summary of some of the validated techniques from the literature to measure user experience in general, and many of these techniques can be easily modified for a specific technology classes such as AI to quantify the experience of the user.

00:43:48.000 --> 00:43:52.000
Which in turn can influence the degree of trust in the system.

00:43:52.000 --> 00:43:58.000
Some of you, for example, may be familiar with the NASA TLX method, which is also listed in this table.

00:43:58.000 --> 00:44:10.000
Now, fairness takes this conversation one a step further in the sense that it describes a situation where a technology such as an AI system is not only trustworthy,

00:44:10.000 --> 00:44:20.000
but also generates equitable outcomes across individuals or groups of users.

00:44:20.000 --> 00:44:32.000
Unlike trust, fairness is more difficult to quantify and generalize. For instance, we don't really know our fairness looks like in many problem domains, even when humans make decisions.

00:44:32.000 --> 00:44:40.000
Also, training AI models with existing data may not be sufficient if universal fairness is the ultimate goal.

00:44:40.000 --> 00:44:45.000
And lastly, as AI systems are becoming more intelligent and human-like,

00:44:45.000 --> 00:44:56.000
they show tendency to fall into the trap of mimicking human behavior and more specifically, biases in making inferences with human generated data.

00:44:56.000 --> 00:45:12.000
An interesting demonstration of the lack of fairness in some of the recently developed generative AI models can be seen in this example, which I took from the latest edition of the AI Index Report by Stanford University. Here are the human prompt given to four different AI models

00:45:12.000 --> 00:45:22.000
was influential person, and as you can see, all four AI models generated human faces that were strikingly similar.

00:45:22.000 --> 00:45:30.000
A more detailed analysis of some of these tools shows that generated AI content in addition to being biased with respect to gender 

00:45:30.000 --> 00:45:45.000
features unfair outcomes along age and race. So it is really critical to take a closer look at AI fairness, particularly when the outcomes have major societal implications such as in the disaster management domain.

00:45:45.000 --> 00:46:02.000
Fairness can be viewed through the lens of the input data and whether it perpetuates social biases the model's internal functions and whether they learn or amplify data biases and the sensitivity of the outcomes to certain population characteristics.

00:46:02.000 --> 00:46:06.000
If you take disaster damage assessment as an example application domain,

00:46:06.000 --> 00:46:17.000
one way of making sure that AI outcomes are fair or at least are not worse than the decisions made by the human expert is to have AI explain how it has achieved those outcomes.

00:46:17.000 --> 00:46:23.000
This is an area that we have been looking into by trying to add explainability to AI,

00:46:23.000 --> 00:46:33.000
and essentially creating models that not only do output damage score for a building impacted by disaster, let's say in accordance with FEMA guidelines 

00:46:33.000 --> 00:46:43.000
or the instructions by other federal or state agencies but also show us what components of that building it has analyzed to arrive at that conclusion.

00:46:43.000 --> 00:47:01.000
This allows us to then refine the model for future iterations as needed to make sure that when the model analyzes these images, it does not rely solely on things that are more prevalent in certain communities that are disproportionately impacted by disasters,

00:47:01.000 --> 00:47:21.000
but keep receiving less assistance to rebuild and recover. Examples of these items include deferred maintenance graffitis on the walls, poor landscaping, and others that may lead to skewed AI outcomes.

00:47:21.000 --> 00:47:33.000
Another obstacle in achieving large-scale fairness is that AI model developers have been quite greedy in consuming the entire universe of existing data to train large and generalizable AI models 

00:47:33.000 --> 00:47:40.000
over the past few years, for instance, AI-enabled large language models have become larger and more complex,

00:47:40.000 --> 00:47:53.000
and been trained using more computing power and more parameters and still many such models fall short of achieving fair outcomes in new contexts and where it is needed the most.

00:47:53.000 --> 00:47:59.000
The literature lists three possible approaches in creating high-performing AI models.

00:47:59.000 --> 00:48:08.000
Using more training data increasing the number of model parameters or adding to the number of computing operations executed during model training.

00:48:08.000 --> 00:48:17.000
To date, the majority of models generated with small data although having more parameters and more computes 

00:48:17.000 --> 00:48:26.000
tend to be slow and expensive, which implies that more data is still a hard requirement for AI model development.

00:48:26.000 --> 00:48:42.000
As such, if we want to achieve fairness at scale for all constituents, we must capture, understand, and incorporate vast amounts of data. However, at current rates, it is expected that by 2028, the typical size of AI training dataset

00:48:42.000 --> 00:48:48.000
will reach the same size as the total estimated stock of public textual data,

00:48:48.000 --> 00:48:59.000
and the extent to which new or generated or unconventional data sources comply with our expectations of fair and equitable AI outcomes remains an open question.

00:48:59.000 --> 00:49:11.000
This is an interesting visualization showing that the size of training data for large language models is rapidly catching up with the amount of available text on the Internet. And according to one expert.

00:49:11.000 --> 00:49:18.000
even 10 times increase in the stock of data only buys you approximately three years of time 

00:49:18.000 --> 00:49:34.000
before we run out of data again. Lastly, I would like to also mention that sometimes our goal of making AI so intelligent and human-like may contradict with our goal of having bias-free and fair AI outcomes.

00:49:34.000 --> 00:49:46.000
After all, human captured data is by definition biased and therefore, using this data to create models that can not only make causal inferences and reasoning,

00:49:46.000 --> 00:49:52.000
but also consider the moral and ethical implications of the generated decisions is not a trivial task.

00:49:52.000 --> 00:50:02.000
Point in case even as humans, we may face moral dilemmas and disaster scenarios for which we may not have a good solution.

00:50:02.000 --> 00:50:09.000
So in summary, I would like to emphasize again the importance of effective and lasting community engagement.

00:50:09.000 --> 00:50:27.000
In collecting AI ready data and in designing, developing, and deploying AI systems in any context, but specifically in the disaster management domain where even a small disparity in the outcome of the model can lead to a large and disproportionate impact

00:50:27.000 --> 00:50:34.000
on certain at-risk and marginalized communities. Thank you for this opportunity.

00:50:34.000 --> 00:50:41.000
And for listening to this short presentation.

00:50:41.000 --> 00:50:49.000
Thank you, Amir. And just a reminder, anybody can put questions in the talk for our speakers.

00:50:49.000 --> 00:50:56.000
Or in the chat for our speakers. And our final speaker for today is Sorrelle Friedler from Haverford College. 

00:50:56.000 --> 00:51:07.000
Who will be talking about how to successfully use AI for science, experimental design, explainability, and avoiding bias. Take it away Sorelle.

00:51:07.000 --> 00:51:16.000
Hi, everyone. It's good to be here with you today. I'm Sorelle Friedler. I'm a professor of computer science at Haverford College 

00:51:16.000 --> 00:51:28.000
and formerly the assistant director for data and democracy in the White House Office of Science and Technology Policy. While I was at OSTP, I was part of the group that developed the AI Bill of Rights,

00:51:28.000 --> 00:51:40.000
and many of the protections that we recommend in the AI Bill of Rights are also actually useful when thinking about AI for science, especially thinking about, of course 

00:51:40.000 --> 00:51:48.000
safety and efficacy, but also avoiding bias and including explainability. So I wanted to talk about those today 

00:51:48.000 --> 00:51:56.000
in the context of a project. That I started working on almost a decade ago, which has since largely wrapped up 

00:51:56.000 --> 00:52:09.000
called the Dark Reactions Project. In this, I was working with colleagues in chemistry to try to predict whether a crystal will form as part of a material science experiment.

00:52:09.000 --> 00:52:27.000
And so when thinking about that, we have a general recipe that is the sort of experimental procedure and when we try to use this to develop a machine learning system, then we also need some sort of outcome label 

00:52:27.000 --> 00:52:49.000
that we can assign. And one of the things that's tricky about doing this in a scientific context is, of course, that we need both positive and negative examples in order for these machine learning systems to effectively learn to model the chemical space. And of course, in scientific publications, we largely like to report on the positive results,

00:52:49.000 --> 00:53:05.000
not the negative ones. And so in order to get those negative results, we went back to historical lab notebooks and recovered that information about negative results, right? Experiments that were tried in the lab and failed.

00:53:05.000 --> 00:53:30.000
And taken together, we were then able to develop a machine learning system, in this case, an SVM, that could fairly accurately predict the outcome of chemical experiments. And we incorporated that into a recommendation system that could help us to recommend future experiments that might also be successful while still being different from the past performed ones.

00:53:30.000 --> 00:53:41.000
One of the questions that my colleagues had for me after we were able to develop this machine learning classifier was very reasonably, you know, okay, so it can perform well.

00:53:41.000 --> 00:53:52.000
What does it know about the science that we don't? And so in order to answer that question, we had to develop a way to get the machine learning model to explain itself.

00:53:52.000 --> 00:54:05.000
In other words, when we're thinking about science, it's not always enough just to have the right answer, right? We want to know why. We want to develop scientific hypotheses. And so our goal really was to learn from the classifier.

00:54:05.000 --> 00:54:29.000
And in order to do this, we used a method that's known as post-hoc interpretability. In other words, we want to be able to examine a black box model after the fact. The strength of this is that it allows us to use the most powerful model possible so that we can model the science as accurately as possible while still hopefully extracting some sort of knowledge from it.

00:54:29.000 --> 00:54:38.000
And specifically, in the context of our first paper, we used a technique that we can think of as modeling the model or developing a proxy model.

00:54:38.000 --> 00:54:57.000
And the way that this works is that we create some sort of interpretable model. So a model where we can actually directly examine it, something like a decision tree, right, or linear regression might be another reasonable way to look at this. And we look at the outputs from the model that we want to examine 

00:54:57.000 --> 00:55:03.000
and we train based on those predicted outputs, but we train a new interpretable model.

00:55:03.000 --> 00:55:14.000
So this is not directly examining the actual model. That made the predictions. It's developing sort of a model of that model that we can then directly examine.

00:55:14.000 --> 00:55:21.000
And so we did that for our SVM and we ended up with this decision tree. Now,

00:55:21.000 --> 00:55:31.000
as you can tell, it's not all that directly interpretable, right? It's quite big. It requires a lot of work to interpret,

00:55:31.000 --> 00:55:38.000
but in this scientific context, right, where we wanted to really understand what was going on, it was worth it 

00:55:38.000 --> 00:55:43.000
to us and to my colleagues in material science to spend the time 

00:55:43.000 --> 00:56:02.000
to really examine this model. And so this is a decision tree that describes the decisions that our SVM was making. And you can see that some of the nodes in it are highlighted, and these are actually the portions of the decision tree that on examination, my colleagues decided might warrant

00:56:02.000 --> 00:56:32.000
further hypothesis testing in the lab, right, so they were able to take these sort of highlighted these three highlighted sections, turn them into formal hypotheses and test and confirm them in the lab using normal experimental methods. And so, you know, that gave us confidence both that the SVM model was acting appropriately on the underlying scientific criteria and also that our interpretability method

00:56:32.000 --> 00:56:49.000
could allow us to appropriately examine these hypotheses. There are other methods that one could use also to try to understand what's going on with a model of this type that are also post-hoc and model agnostic, right, meaning they can work on any

00:56:49.000 --> 00:56:56.000
black box or potentially complex model. And one of the common ways of doing this is auditing for future influence.

00:56:56.000 --> 00:57:02.000
The basic idea here is that we want to quantify the relative importance of features to a black box model.

00:57:02.000 --> 00:57:15.000
So that allows us to look at these complex black box models and to say, you know, in a quantified way comparatively, how important was this particular feature on the outcome 

00:57:15.000 --> 00:57:32.000
of the model. And there are a number of packages now that can be used to do this. One of the really common ones is known as SHAP, and it's used all over the place partially because they developed really nice code. It's really easy to use. It creates nice graphics.

00:57:32.000 --> 00:57:52.000
Unfortunately, it has a lot of drawbacks. So there's been subsequent research showing that it has no real clear or meaningful interpretation in a future importance context in the context of worrying about discrimination that can actually be adversarially hidden from

00:57:52.000 --> 00:58:22.000
SHAP. It also has been shown not to be useful for debugging, which is really important in a scientific context, right? We want both a clear and meaningful interpretation and the ability to see, you know, just at its basic, does this model make sense in a scientific context? It also doesn't allow for recourse, which is a property where when a feature is changed, you expect the outcome to be able to change, which is also, I would say, useful in a scientific context, right. We want to know, okay, if this feature was really important to a model, does that mean

00:58:27.000 --> 00:58:33.000
that if I change it, I might be able to get a better outcome or a different outcome. Let's say if it was a predicted failing experiment 

00:58:33.000 --> 00:58:45.000
could I shift the most important feature so that it would now succeed. And it turns out that this method for feature attribution does not allow you to make that claim.

00:58:45.000 --> 00:59:14.000
On the other hand, if what you're looking for is some sort of feature attribution methods, as long as you're really clear about what your goals are, there are now other feature attribution methods out there that can give you some of the confidence. There's one known as Weight of Evidence that has been shown to be useful for debugging. In an anti-discrimination context there's a number of future importance methods that have been developed to try to address that. And I have a recent

00:59:14.000 --> 00:59:32.000
paper with ongoing work that can handle this in a recourse context, right. Where we can give a guarantee that if you change the feature that we identify as important in this context, that it's more likely to give you a new outcome,

00:59:32.000 --> 00:59:38.000
right, which I think may also be useful in a scientific context.

00:59:38.000 --> 01:00:00.000
Another of the key concerns that we raise in the AI Bill of Rights and that has been raised throughout the literature at this point is a concern about bias. And here the concern is often about what happens when you train models that have to do with people.

01:00:00.000 --> 01:00:19.000
And, in these contexts, we often end up with a situation where the resulting model is discriminatory, not necessarily because the person who created the model wanted it to be discriminatory, but often because the underlying societal data that it's trained on essentially has baked in discrimination.

01:00:19.000 --> 01:00:37.000
So you might imagine that training on historical data, you essentially repeat the problems of the past. So for example, a company that trains on information about software engineers at their company, but which has historically hired predominantly men 

01:00:37.000 --> 01:00:57.000
found that when they tried to create a resume screening algorithm based on that data that it screened out women software engineers resumes, right, because these algorithms are just trying to replicate patterns of the past. It turns out that we can show that we have that same problem in a scientific context.

01:00:57.000 --> 01:01:13.000
Right. So, human scientists also have these types of anthropogenic biases. So people are more likely, for example, to choose reagents that are on their shelf.

01:01:13.000 --> 01:01:21.000
Currently, right, and so they might choose the same ingredients to an experiment over and over again and the same experimental conditions, right.

01:01:21.000 --> 01:01:41.000
So they might not vary the choices in this sort of experimental recipe as much as they would under, say, random choices. And it turns out that you know, while while we might think or hope that these instincts are in fact

01:01:41.000 --> 01:02:06.000
you know the result of hard-won expertise. In some cases, they're not, and it turns out that these sort of, you know, anthropogenic biases lead to outcomes that are less successful or really not more successful than randomly sampled choices, right. They have effectively the same outcome

01:02:06.000 --> 01:02:13.000
chance as if we were to randomly sample from a carefully designed, you know, random space.

01:02:13.000 --> 01:02:34.000
And what's more, I think this is the more relevant piece when it comes to thinking about creating models, this randomly sampled data is more effective when we use it as training data, right. So if part of the goal in conducting a suite of experiments is

01:02:34.000 --> 01:02:43.000
to both, right, have successful experiments, but also to generate data that might be useful when training models, then it's much more effective 

01:02:43.000 --> 01:02:57.000
to carefully create a random sample that will have potentially the same likelihood of success, but additionally create 

01:02:57.000 --> 01:03:01.000
more effective models.

01:03:01.000 --> 01:03:22.000
And so, in summation it is important to create the experimental mental conditions carefully, right, so that we avoid these anthropogenic biases so that we have both successes and failures in our data. But also it's important to use methods that will allow us to

01:03:22.000 --> 01:03:34.000
avoid bias and to explain the results in a way so that hopefully we can learn from the models how to do better science.

01:03:34.000 --> 01:03:39.000
Thank you so much.

01:03:39.000 --> 01:03:42.000
Thank you, Sorelle, and thank you to all of our speakers.

01:03:42.000 --> 01:03:49.000
Let's begin the discussion session. So if you have any questions, you can put them in the chat.

01:03:49.000 --> 01:03:55.000
And we already had some pop up.

01:03:55.000 --> 01:04:09.000
For example, there's a question for Carrie Ann, which I think is a really good one to start off the discussion session because machine learning is all about big data But for big data.

01:04:09.000 --> 01:04:17.000
What is big data? Like how big is large in terms of the data set or a parameter space?

01:04:17.000 --> 01:04:21.000
And does it depend on the problem?

01:04:21.000 --> 01:04:38.000
Yeah, I guess. Big, I mean, it depends a lot on the particular task and the model. So the way that I think about it is the machine learning model is learning mapping from whatever sort of inputs you give your model to the output thing you want to predict.

01:04:38.000 --> 01:04:41.000
And so there's going to be some complexity to that mapping.

01:04:41.000 --> 01:04:50.000
And some of that's going to depend on the specific task and also the specific architecture you're choosing. If you're choosing a simple model, like say a linear regression model.

01:04:50.000 --> 01:05:07.000
Then that's, you know, you have a small number of parameters to learn. You might need less data for that versus a really complicated, like deep neural network is going to have a lot of parameters it's going to be able to learn a much more complex model. So the more complex your task is, the more complex the machine learning model you need, and typically the more data.

01:05:07.000 --> 01:05:22.000
But there's also ways of, as I said, bringing in other forms of data and information. So if there's information from related tasks, like if you could take one of these foundation models, which is able to kind of extract features from related data sets, then you might be able to take those features

01:05:22.000 --> 01:05:39.000
Some of what that model has learned and then leverage it for your task with less data. So it really depends on Or for instance, if you're able to use information from physics and embed some sort of physics into your model. So maybe rather than trying to relearn all the physics, you're learning a few physical parameters.

01:05:39.000 --> 01:05:47.000
Then you might be able to learn with less data because you sort of reduce the size of your search space. So it really depends also on what other sort of information you're bringing in.

01:05:47.000 --> 01:05:58.000
And maybe the other speakers would like to add, I'm sure they have something that they can contribute as well.

01:05:58.000 --> 01:06:11.000
I was the one that asked the question and I was interested. Hi, I'm Danny from UC Berkeley and I do inside stuff with landslides. And I was like, how many landslides do I need to map before I can have a big enough data set to use it?

01:06:11.000 --> 01:06:28.000
To like train the mapping aspect, which is, I think, pretty simple, right? Where it's like some pixels are moving faster than the ones around it. So how many do I need to like manually map before I can like start a machine learning project?

01:06:28.000 --> 01:06:34.000
I mean, I would say to that, one thing that is kind of nice about machine learning is at least its structure is very simple.

01:06:34.000 --> 01:06:43.000
And so if you wanted, I mean, we could talk about this too, if you'd like, like it doesn't take all that much time to set up a machine learning model to try to train on data.

01:06:43.000 --> 01:06:53.000
Depending on the data that you want to look at so like if you have the time, you could always just give it a shot and see you know given the data that you have, what happens if you hit it with machine learning, right? Does it do

01:06:53.000 --> 01:06:57.000
What you want to do or whatever right like i don't know.

01:06:57.000 --> 01:06:59.000
I'm going to email you, Cody. I'll talk to you later.

01:06:59.000 --> 01:07:04.000
Do it, please.

01:07:04.000 --> 01:07:18.000
I mean, so I had a question for Cody. So at least in recent news, there's the deep seek approach, right? So if people haven't a group has made this open source ai tool that's sort of equivalent to chat GPT

01:07:18.000 --> 01:07:25.000
But the real breakthrough is at least the creator's claim that it took only a fraction of the resources to train the model.

01:07:25.000 --> 01:07:31.000
I mean, so does that help with the energy consumption? Are there takeaways that we can use in geoscience?

01:07:31.000 --> 01:07:39.000
I mean, yeah, that's actually very interesting. I hadn't heard about that because that was kind of what I was seeing with the carbon footprint problem is like you have some models that contained similar accuracy across, you know.

01:07:39.000 --> 01:07:54.000
Different very like levels of required energy If you're able to achieve, and this comes down to like, you know, architecture and like what we're actually building these models and like how we should be building them. If it turns out that there's a more efficient architecture for something like ChatGPT,

01:07:54.000 --> 01:08:04.000
Then the kind of solution to the environmental problem comes down to finding those more efficient I'll have to look into that. That is super cool.

01:08:04.000 --> 01:08:11.000
I think it's pretty recent news at least you know a lot of the stocks took a tumble the past few days.

01:08:11.000 --> 01:08:12.000
Oh.

01:08:12.000 --> 01:08:20.000
Over the news, yeah. But yeah, so I guess, I mean, kind of related to that for Amir, right, talking about, you were talking about, hey, we're running out of data to train on.

01:08:20.000 --> 01:08:35.000
But if we create these new frameworks, new methods of approaching these problems, right? That could also be a solution too, right?

01:08:35.000 --> 01:08:46.000
On principle, yes. So you have like three, like I said in my talk, there are three approaches to make a model stronger and high performing and One of them is obviously using more data.

01:08:46.000 --> 01:08:52.000
And to date, it's been shown that the models trained on more data are essentially more robust.

01:08:52.000 --> 01:09:00.000
And more generalizable. Now, you can also argue that you can increase the complexity of the model or add more computes at training.

01:09:00.000 --> 01:09:14.000
And those are also approaches that have been tested by different researchers. I haven't studied that much about deep seek. The news just came out yesterday or like two days ago. So it remains to be seen what exactly is happening there.

01:09:14.000 --> 01:09:21.000
But based on the available literature. Many people argue that if you want to have generalizable models.

01:09:21.000 --> 01:09:33.000
Particularly in the context of having fair and trustworthy models, having access to a large amount of data might be the ultimate solution.

01:09:33.000 --> 01:09:34.000
I see, I think Ben's hand is up. Do you have a question, Ben?

01:09:34.000 --> 01:09:36.000
Thank you.

01:09:36.000 --> 01:09:47.000
Yeah, hi. Thanks, everybody, for the great talks. I think for Amir or really for anybody, what's the estimate of the amount of non-digital data that are out there?

01:09:47.000 --> 01:09:53.000
Yeah. Is it significant? It's certainly interesting.

01:09:53.000 --> 01:09:54.000
But is it significant?

01:09:54.000 --> 01:10:05.000
I was… Yeah, that's a great question. Thank you for asking that. I was actually debating whether or not to add one more slide when I was recording my talk, but I eventually decided not to add that.

01:10:05.000 --> 01:10:23.000
There are some estimates. There is this AI research company called Epoch And back in 2023, they had a study out there that essentially projected the the time by which machine learning trained data will be exhausted.

01:10:23.000 --> 01:10:38.000
With a 90% confidence. What I presented was obviously the language stock and they have also another estimate for the image stock which puts that exhaustion time somewhere around 2038.

01:10:38.000 --> 01:10:48.000
And that's what current estimates and then you know whether we are relying solely on people capture data or whether AI generated images will also be added eventually to this stock.

01:10:48.000 --> 01:11:02.000
But again, we have to keep in mind that if AI models are biased and generated images using AI might also have implicit biases in them. But their estimate essentially says another 12 to 13 days before we run out of

01:11:02.000 --> 01:11:10.000
Image data which is the foundation for many of the computer vision and machine learning models.

01:11:10.000 --> 01:11:24.000
Well, keeping with the theme of pets then, or Sarah's favorite cats, I mean, this is why we want to have like ubiquitous cat cams Because then we're going to learn the world from Kat's images or from their own viewpoint right so

01:11:24.000 --> 01:11:36.000
Maybe there's hope after all, right, sir?

01:11:36.000 --> 01:11:44.000
And one thing I want to add is that also is These estimates are primarily based on publicly available data for general machine learning applications.

01:11:44.000 --> 01:12:08.000
I think one of the speakers was talking about the numerical weather prediction models and going from low resolution climate data to high resolution climate data, most of that data remains proprietary so given you know as we speak we may not have enough data in the next few years to train our machine learning models for those specific applications. But cat and dog images, I'm sure we can

01:12:08.000 --> 01:12:11.000
Figure out the way to increase the number of those images.

01:12:11.000 --> 01:12:20.000
No, but I mean to understand the world from the way that a cat perceives it, which we currently don't really have a way of doing.

01:12:20.000 --> 01:12:21.000
Sure. Yes.

01:12:21.000 --> 01:12:29.000
There's certainly knowledge there. There's this whole Animal Kingdom of knowledge that we don't have real access to.

01:12:29.000 --> 01:12:30.000
Right. Yeah. Cool.

01:12:30.000 --> 01:12:37.000
Absolutely. I'm a fan of those that are like nature programs where they put cameras on migrating birds and you can see the bird's eye view when they I think that's a great idea.

01:12:37.000 --> 01:12:46.000
Right, right, right, right, right.

01:12:46.000 --> 01:12:47.000
Well, I have a question for

01:12:47.000 --> 01:12:52.000
Maybe learn about earthquake prediction that way too.

01:12:52.000 --> 01:12:59.000
I have a question for Sorrel. You mentioned the importance of publishing both successes and failures in data.

01:12:59.000 --> 01:13:05.000
However, there's not much of an infrastructure for publishing failures, unfortunately.

01:13:05.000 --> 01:13:10.000
So do you have any thoughts on how a scientist might be able to achieve that goal?

01:13:10.000 --> 01:13:20.000
Yeah, I mean, I agree it's hard. So, you know, I guess sort of following on the previous conversation, I think I'm more of a skeptic that what we need here is just like.

01:13:20.000 --> 01:13:40.000
More data. I think what we need is better data. And I don't know that we need a huge amount of it, right? So the project that I was talking about in material science, there were a couple of different sort of specific domains that we targeted, but in each of them, I would say we had

01:13:40.000 --> 01:13:53.000
Between maybe three and 5,000 instances, right? Now those were experiments, each of which took days, right? So we're really sort of drawing on my colleagues lab notebooks going back a decade.

01:13:53.000 --> 01:14:12.000
So it's a lot of data in that sense. But in a machine learning sense, when we're talking about data set sizes that are in the thousands. That's you know really small data, but we were still able to make it quite effective, right? And so, you know, what I would say is like.

01:14:12.000 --> 01:14:17.000
What are those other data sets out there that people have been collecting and sort of carefully recording in the lab?

01:14:17.000 --> 01:14:32.000
Or, you know, could be carefully recording in the lab and how could those be sort of turned into digitized data in such a way so that they could then build on them and perhaps that's then the thing that they publish, right?

01:14:32.000 --> 01:14:52.000
All of our data is, of course, published with our papers, including the negative instances, so that it can be helpful to other folks training or trying to reproduce the results. And so, yeah, while that's not sort of publishing directly the negative results, it does at least include that data.

01:14:52.000 --> 01:14:56.000
I think Tim has his hand up. You have a question, Tim?

01:14:56.000 --> 01:15:00.000
Yeah, so I just want to say thank you for everyone here in this great talk.

01:15:00.000 --> 01:15:08.000
I kind of a comment and a question. So it seems like Cody and Carrie Ann are kind of talking about AI ML for doing kind of pure like science discovery kind of thing.

01:15:08.000 --> 01:15:27.000
And then Amir and Sorrell what's the interaction between AI and kind of humans and applying the science to you know problems that are important. So I'm kind of wondering, should these two kind of like AI for science discovery, AI for interacting with humans, should they be the same thing or do we need different methods and procedures for doing these two different kind of

01:15:27.000 --> 01:15:34.000
Styles of research. Especially kind of open for anyone.

01:15:34.000 --> 01:15:41.000
So I guess I would… jump in quickly and say that, you know, I think that you can't have, at least in my experience.

01:15:41.000 --> 01:15:47.000
I don't think that you can have the AI for science without also including the humans.

01:15:47.000 --> 01:16:00.000
And that's sort of actually how I ended up getting involved in AI Explainability and sort of AI human in the loop work was essentially because I was trying to just do AI for science and my scientific colleagues were like.

01:16:00.000 --> 01:16:08.000
But what's going on in there, right? I think a good question, right? And one that we really couldn't get past.

01:16:08.000 --> 01:16:14.000
Without actually trying to directly tackle.

01:16:14.000 --> 01:16:21.000
I think from my perspective, if you talk about trustworthy AI, Just like the framework I presented showed.

01:16:21.000 --> 01:16:40.000
You need to have human interaction with the system. So you can have the highest competent model that meets your expectations in terms of accuracy and reliability but If a human user cannot interact with that model or doesn't have a pleasant experience or engaged experience with the model

01:16:40.000 --> 01:16:57.000
It's very likely they're not going to trust that model. And the model remains you know becomes obsolete because when the humans don't use it, you're not going to collect your data. You're not going to be able to fine tune the model based on new experiences. And therefore, you're not going to be able to generalize it to unseen conditions that you are not

01:16:57.000 --> 01:17:18.000
Considering during the training. So while I'm an advocate of having highly competent and accurate models, I'm also advocating for bringing these models to the communities and let them letting them experience with the models and show us cases where the models fail because to me, I think someone mentioned in the chat box as well

01:17:18.000 --> 01:17:44.000
Those examples of model failures are more informative than just the success examples because success examples are extensions of what my training data can tell me about failures are probably cases where the training data lacked the ability to provide insight into those cases.

01:17:44.000 --> 01:17:51.000
Oh, Carrie Ann, I have a question for you. We've heard a lot about biased data in a societal context.

01:17:51.000 --> 01:17:59.000
But could you give us an example of data bias in the context of earthquake hazards and how to avoid it?

01:17:59.000 --> 01:18:14.000
Yeah, I guess I can do my best. I think one of the places in the sciences where we see biases, like there's kind of a bias towards where we have instruments, what's easy to collect what you know it's there's certain

01:18:14.000 --> 01:18:19.000
Like we have more data coming from countries that have a higher budget for instruments, right?

01:18:19.000 --> 01:18:39.000
And so for some kinds of data like remote sensing data, maybe we have global coverage, but for other cases where we need to have, you know, or have experience where you need someone to go into the field, then those are going to be biased towards locations that are going to be easy to get access to, for instance, that they're going to be in places that are near roads or have better infrastructure or have

01:18:39.000 --> 01:19:03.000
Safer to visit or just more accessible. So those kind of biases are always going to be baked in in geoscience data, but a lot of times it's going to be about what's easy to collect. It also might be if we're talking about extreme events, one of the challenges we have, some of the projects I've been working on with people here, we're interested in what's going to happen with future climate. And what people are really interested in are like the extreme cases, like what's going to happen if a big storm comes through.

01:19:03.000 --> 01:19:10.000
Not what's going to happen on kind of the average day, but what are going to happen in these extreme cases. And so we don't have as much data on extreme cases as well.

01:19:10.000 --> 01:19:15.000
And they might not be representative or representative of the kind of extremes we'll see in the future. So that's another.

01:19:15.000 --> 01:19:33.000
Challenge. And so I think in terms of addressing it, I mean, I think we have to be aware of these things when we're doing our modeling. Sometimes you can't just like magic data into existence that you wish you had. But if you're thinking about what data you want to collect using things like synthetic data could be potentially a solution, simulated data.

01:19:33.000 --> 01:19:50.000
And trying to make sure that you're trying to bring in diverse data sources is something that you try to do. But sometimes you're just limited on what you can do. And in that case, it's something that you want to be aware of in the analysis and then doing like testing to make sure that it's actually going to work in these cases. But I don't really have like a

01:19:50.000 --> 01:20:06.000
A good solution for you, I guess. But I think the first step, it's thinking about the data, where it's coming from, where it's collected and what those biases might be and being aware of it is certainly important.

01:20:06.000 --> 01:20:12.000
So sort of related to that, Sorrel, you talked about that humans are biased in our experimental choices.

01:20:12.000 --> 01:20:26.000
But that seems almost kind of counter to the idea that domain knowledge is useful. I mean, so how do we approach this? Do we make our ideal domain knowledge model and then just test random things too? Or how do we kind of weigh those two

01:20:26.000 --> 01:20:28.000
Factors.

01:20:28.000 --> 01:20:58.000
Yeah, I mean, so I did not mean to imply that domain knowledge wasn't important. It was in fact critical, right? So like one of the One of the common biases that we found in our specific work was that in the literature, there was sort of a bias towards some specific reagents. And those are the ones where I was saying, you know, okay, if you randomly choose across these, you actually get the same success. It's just what people happen to have on their shelf. So that's what they reach for, right? Makes sense.

01:21:00.000 --> 01:21:11.000
On the other hand, we also found biases based on being able to actually dissolve things in water.

01:21:11.000 --> 01:21:28.000
And like if you increased the amount enough that it didn't dissolve in water, sure, that was like a part of the experimental space that the material scientists had not explored but like there was a reason, right? You know, there were some some actual sort of um

01:21:28.000 --> 01:21:45.000
Physical and chemical boundaries that you could not cross and that our model did not know about. Right. And so occasionally, as we were developing the recommendation system, the model would sort of suggest to my chemistry colleagues that they perform experiments. They were like, this is nonsensical. Like I don't

01:21:45.000 --> 01:22:05.000
I don't need to go to the lab to try this. I know it doesn't work, right? And so, you know, I think as sort of other speakers have also mentioned, right, like being able to put those sort of physical bounds into the model is really useful, right? Using that domain knowledge to sort of put in

01:22:05.000 --> 01:22:13.000
Put in those bounds on the experimental space where they're known is a great way to augment them.

01:22:13.000 --> 01:22:15.000
Do you mind if I say something on that as well?

01:22:15.000 --> 01:22:16.000
Please do.

01:22:16.000 --> 01:22:29.000
I just think it's a very interesting question because when you're especially in modeling you have to make assumptions about the system, right? You kind of have to create this perception of how the universe is going to behave. And one thing I think that's really interesting is

01:22:29.000 --> 01:22:37.000
The kind of way in which you can probe those assumptions, right? It's like if you're using a machine learning model and the data is disagreeing with the kind of model assumptions that you come up with.

01:22:37.000 --> 01:22:48.000
I think it's worthwhile to go back and ask yourself like what assumptions here are realistic? Which ones are causing the system to behave in a way that maybe isn't consistent with what we're actually observing.

01:22:48.000 --> 01:23:04.000
And that's just something I think is really interesting about kind of these more data-driven approaches is kind of the it's like a closer merging of the empirical with the kind of theoretical.

01:23:04.000 --> 01:23:18.000
Yeah, Cody, kind of on that note, what steps can we take to ensure that models in the earth scientists are interpretable and transparent for non-technical stakeholders.

01:23:18.000 --> 01:23:27.000
So, I mean, interpretability is always kind of a tricky one um Especially, you know, for depending on the architecture and the machine learning that you're kind of doing.

01:23:27.000 --> 01:23:33.000
One thing, at least in the field that we work in fairly closely is in the physics informed aspect of it.

01:23:33.000 --> 01:23:43.000
There is a level of interpretability that you get that a lot of other kind of approaches don't in that it's you can directly check whether it's satisfying the visible laws that you told to kind of impose.

01:23:43.000 --> 01:23:49.000
So at the very least, at the end of the day, even if you don't know exactly what's happening under the hood, you can at least say, hey, it's satisfying my boundaries.

01:23:49.000 --> 01:23:55.000
Satisfying my initial conditions. And it's satisfying my PDE that I've told to satisfy.

01:23:55.000 --> 01:24:02.000
And you can quantify how closely it satisfies that. I think at least in the realm of modeling that I work with.

01:24:02.000 --> 01:24:20.000
There is a little bit more interpretability in terms of the physical modeling But if you want to probe kind of what's under the hood, I'm not quite ready to say one way or another how to go about doing that.

01:24:20.000 --> 01:24:33.000
Does anybody else have any ideas on that question?

01:24:33.000 --> 01:24:55.000
I guess I would just quickly jump in and say I think it depends on exactly what the goal is for those non-technical stakeholders, right? You know, in some cases, they're just going to want to know like what part of my data did you use, right? You know, or is this uh like is this

01:24:55.000 --> 01:25:11.000
Are these results relevant to my house? In other cases, they might want to know much more detail, right? And so I think it's useful to sort of think about as in everything exactly who you're communicating to and what sort of

01:25:11.000 --> 01:25:15.000
Pieces of information they might care about.

01:25:15.000 --> 01:25:29.000
Yeah, I guess I can jump in a little bit. So in my group, we're working a little bit on trying to develop some explainable AI techniques and we're trying to build techniques where we kind of build it into the architecture so that they're a little more intrinsically interpretable but with

01:25:29.000 --> 01:25:41.000
Any method, whether it's this post hoc techniques or building the sort of models to be designed to be easier to interpret. Like you're only going to get a partial view. And so a lot of times it comes back to kind of as the other speakers have said like

01:25:41.000 --> 01:25:53.000
What are people looking for to understand about the model? Because you're only going to be able to understand it like certain aspects. You're not going to be able to like if you have a huge model with millions of parameters. So you're not going to be able to understand necessarily every little

01:25:53.000 --> 01:26:17.000
Thing about it. So sometimes, and in some cases, like in this collaboration I have with climate scientists who are building these emulators of these ice sheet models Sometimes what they're interested in wasn't actually interpretability as much as like, well, how confident are we in the model prediction? Is it making decisions using in a way that makes sense to us? Like if we leave out this variable that we don't think matters.

01:26:17.000 --> 01:26:31.000
And build the model, is it still going to kind of give the same results if we add these different variables in, which variables are getting more weight in the decision and that sort of thing so that it kind of aligns with their own perception of what they would expect the model to be doing to help get confidence and then having those kind of

01:26:31.000 --> 01:26:44.000
Uncertainty or error bars along with the prediction so that they say, okay, we're not just getting a point estimate. So I think some of it is like talking to the people who would be using your models And determining what is it that they actually, what information would they actually like

01:26:44.000 --> 01:26:47.000
Along with the predictions or what sort of analysis should we do?

01:26:47.000 --> 01:27:00.000
To help them feel more confident, but it's not going to be across the board, like the same for every application or every stakeholder, I think.

01:27:00.000 --> 01:27:02.000
Sarah, did you want to jump in here?

01:27:02.000 --> 01:27:11.000
I just wanted to ask a question sort of following up on the conversation about building.

01:27:11.000 --> 01:27:29.000
Reliability and how the human interface Which is that we struggle even with the physics-based models in terms of making them relatable and understandable and trustworthy to the public, in part because earthquakes happen so infrequently and because people have such little lived experience with earthquakes.

01:27:29.000 --> 01:27:42.000
And so I was, and also I think Ahmed is asking sort of a related question in the chat about, you know, well, maybe data we're not going to run our data with earthquakes because we don't have very many data, but also like we don't have a lot of training data with earthquakes.

01:27:42.000 --> 01:27:59.000
And so I was wondering if I was wondering if I mean, or anyone has given thought to how to do outreach, build trust when you're talking about something that people don't have a lot of experience with and don't have a chance to interact with very often.

01:27:59.000 --> 01:28:05.000
Yes, thank you for asking that question. Actually, I was typing a response for Ahmed's question in the chat box.

01:28:05.000 --> 01:28:21.000
One thing I want to mention, you mentioned earthquakes. I would also add to that that We are witnessing, for example, floods in many parts of the country that historically have not experience floods. So it's not just about you know these acute events like earthquakes, but it's also about

01:28:21.000 --> 01:28:30.000
Events that we are very familiar with because we are working in that domain but the majority of the society hasn't seen one or hasn't experienced one.

01:28:30.000 --> 01:28:54.000
And one thing that's one thing that we have been doing in my lab is using essentially immersive simulation techniques and then having people come to the lab and then experiencing in virtual reality how a flood could look like in their neighborhoods not just you know a typical flood scenario somewhere random in a mountain but also like in the intersection of main and church

01:28:54.000 --> 01:29:00.000
In a particular town that they live. And with the buildings surrounding that they're familiar with.

01:29:00.000 --> 01:29:15.000
And then what we have done is we try to see if their perception of the risk of the flood and how they navigate the way out of that flood situation matches what an intelligent model prescribes or recommends. And if it doesn't.

01:29:15.000 --> 01:29:30.000
Then what why is that? Is it because they don't trust the model? Is it because they don't understand the model? Is it because they have their own way of thinking that we cannot really capture. And this is one way of doing it essentially

01:29:30.000 --> 01:29:41.000
Creating realistic immersive simulations of events that may not be that frequent or of events that may not have happened historically in a particular location.

01:29:41.000 --> 01:29:57.000
And also, just to kind of recap what I wrote for Ahmed's question in the chat box, another challenge we have is that again in the domain of disasters, we are moving more and more towards those tails of the distribution, meaning that the newer events are becoming more extreme.

01:29:57.000 --> 01:30:07.000
And are becoming a phenomenon that we haven't seen before. So the majority of the models, for example, for damage assessment or for the extent of the impact.

01:30:07.000 --> 01:30:22.000
That have been trained on historical data may not really work in a few years, in five to 10 years, because the nature and the behavior and the pattern of these disasters are exponentially changing and just point in case that the recent fires in Southern California compared to

01:30:22.000 --> 01:30:34.000
What people experienced some few years ago the scale is different, but also the pattern of destruction is different. So if you're talking about using AI for building damage assessment, for example.

01:30:34.000 --> 01:30:44.000
Then these models may not really be performing as expected if trained only on historical data.

01:30:44.000 --> 01:30:53.000
Thanks, Sarah. Rachel, would you like to ask your uh stress drop question, which I'm assuming is about stress drop

01:30:53.000 --> 01:31:15.000
Well, about stress drop and other things yes i was also thinking of fine art fault inversions and many of the other things. Basically, the question I put in the chat is that there are a lot of controversies and uncertainties in our science where different people start off with the same data set and end up with

01:31:15.000 --> 01:31:45.000
Different conclusions with different implications. And so, you know, clearly there's probably a role in here for machine learning But if you took interpretation A and their input data and their results as the training set, presumably you would end up reaching your machine learning model would come up with something consistent with their interpretation. But if you took a study which had started off with a completely different

01:31:47.000 --> 01:32:06.000
Result, then they would put that into the machine learning as the training data we presumably get their preferred interpretation So I was as a complete non-machine learning person Is it better to, do you think in these cases to think about

01:32:06.000 --> 01:32:18.000
Unsupervised machine learning where you just give it the data and tell it somehow to come up with something interesting or is it better to give it you know.

01:32:18.000 --> 01:32:26.000
Training data that includes both or multiple interpretations, multiple measurements that disagree and say.

01:32:26.000 --> 01:32:37.000
What do you think to the model?

01:32:37.000 --> 01:32:42.000
I mean, my short answer, I don't know if you guys can hear me.

01:32:42.000 --> 01:32:54.000
My short answer is that it really depends on the nature of the task you're asking the model to do. So unsupervised machine learning Algorithms while they're you know faster because they're just using the data and trying to classify may not be

01:32:54.000 --> 01:33:11.000
Performing as expected if your goal is to get more insight about what is going on with your data. They just classify your Essentially, in most cases, they classify or cluster your data into different categories without really providing an explanation as to why

01:33:11.000 --> 01:33:24.000
And then if your goal is to kind of see similar patterns, maybe unsupervised learning is a better approach. So just to recap, I think it really depends on the nature of the task.

01:33:24.000 --> 01:33:34.000
I guess I would also say that You know, when you're setting up the data to prepare it for one of these machine learning algorithms.

01:33:34.000 --> 01:33:41.000
For most of them, they require some sort of outcome label, right? So, you know.

01:33:41.000 --> 01:33:56.000
While you are able to take sort of a full set of experiments or other observations, right? And hopefully derive some a higher level model mapping between them.

01:33:56.000 --> 01:34:07.000
That doesn't sort of absolve you of the responsibility for writing down the label in the first place. You have to set up the problem and identify like this is the outcome.

01:34:07.000 --> 01:34:16.000
That I'm trying to get the model to predict. And so if that is the site of disagreement, the model is not really going to solve that for you.

01:34:16.000 --> 01:34:25.000
Right. You know, it'll allow you to potentially look at the different possibilities for how you map between the data and that outcome.

01:34:25.000 --> 01:34:41.000
But it's the outcome itself that's that question, then I think that's not going to resolve it for you.

01:34:41.000 --> 01:34:47.000
Great. And then I think we'll ask just a question of all four speakers.

01:34:47.000 --> 01:35:01.000
If you are interested in answering it. The question is, how can we ensure that ML tools developed for earth science in our research are sustainable and usable in the long term?

01:35:01.000 --> 01:35:05.000
So carry on why don't we start with you?

01:35:05.000 --> 01:35:25.000
Yeah, so I guess… some of the things that are helpful I think would be to make the code available, like share make sure that the code is out there, that the data that you've used, if it can't be shared, some of the data sets are large, so it's hard to share, but maybe the code that's used to process it.

01:35:25.000 --> 01:35:35.000
The documentation is all there so that other people can build on the work. I think that's I'm going to also help kind of the field move forward so everyone doesn't have to kind of reinvent the wheel every time.

01:35:35.000 --> 01:35:51.000
And I think there's been a lot of progress on that in terms of since I started working on machine learning and geophysics, it was not very common for a data set and code to be shared. Now there's a lot of groups who are sharing code that they're using, that they're putting out packages, that they're sharing data sets

01:35:51.000 --> 01:36:00.000
And I think having all those things be well documented is going to be helpful so that other people can build on it. I don't know if other people want to add.

01:36:00.000 --> 01:36:10.000
I just… along those same lines i think We are very good in advertising what our models can do.

01:36:10.000 --> 01:36:18.000
We are good marketing people, not us individuals, but as a society when you buy something they always tell you what this can do for you.

01:36:18.000 --> 01:36:31.000
But they rarely tell you what this cannot do for you. So you have to really look at those footnotes or asterisks and kind of zoom in and see what's going on so i think part of documenting the model should be also to illustrate

01:36:31.000 --> 01:36:52.000
Failure cases and the limitations of those corner cases where the model stops performed and that I think adds more to the transparency aspects of the model and therefore people are more will be more inclined to use it for the actual use cases, successful use cases and not trying to kind of push the boundaries and

01:36:52.000 --> 01:36:56.000
Of using it in cases where the outcomes might be consequential.

01:36:56.000 --> 01:37:05.000
In a negative way for the society.

01:37:05.000 --> 01:37:21.000
To answer, I think… something if I think that if we're going to be using machine learning models, I think Transparency is very important in that we're reporting kind of what architecture we're using, what size networks we're going with. But I think even more important

01:37:21.000 --> 01:37:32.000
Reporting on equivalent co2 i think would be really nice And how long the model took to train, basically just being open and honest about the resources that were required to train your model.

01:37:32.000 --> 01:37:43.000
And this is in part because it's good to be out in the open, but also lets us see that when we do make an improvement to models, if we do come up with an architecture that's more efficient for whatever reason, we can measure

01:37:43.000 --> 01:37:59.000
How much more efficient you know how like what gains we actually got from putting in more effort And so I think in terms of sustainability trying to be careful about using our compute resources. We're not Google. We're not going to use the same level of energy as they are but

01:37:59.000 --> 01:38:06.000
Still for developing models, it's good for the environment, but also it allows you to solve bigger and bigger problems. The more efficient your model is.

01:38:06.000 --> 01:38:15.000
Kind of the more you can do with it if you try to scale it up. So depending on how you

01:38:15.000 --> 01:38:21.000
So I agree with other speakers, transparency in all things is key.

01:38:21.000 --> 01:38:32.000
I would really encourage the community to think about ways to share out both the data, but also the model itself.

01:38:32.000 --> 01:38:46.000
There are now platforms like Hugging Face that make that fairly straightforward so that other

01:38:46.000 --> 01:38:48.000
I think we have a… technical issue.

01:38:48.000 --> 01:38:54.000
Don't have to go through the process of just that, you know, not all researchers necessarily have access to.

01:38:54.000 --> 01:38:55.000
Oh, did you get cut off for me? Yes. Can you hear me?

01:38:55.000 --> 01:38:57.000
No, you're good now. Yeah, you're good.

01:38:57.000 --> 01:39:14.000
Okay. So, you know, so I would just encourage you all to think about ways to make the models directly available and also ways to make the underlying training data available, including with documentation. And I would also say including in sort of

01:39:14.000 --> 01:39:24.000
Flat formats like CSV, right? I know it can seem nice to like put it in a fancy database, but that can actually make it really hard to share it across platforms and with different groups.

01:39:24.000 --> 01:39:32.000
So sometimes sort of the simple version and the simple format as much as possible is good.

01:39:32.000 --> 01:39:36.000
I should say one more thing on that? Because I think a lot of people pointed out making the models available.

01:39:36.000 --> 01:39:48.000
This is just something I'm interested in, I like doing, I like developing open source code. So if you're ever interested and you want to kind of get machine learning spun up for your project or you're curious about where you would start like

01:39:48.000 --> 01:39:54.000
Feel free to reach out to me. This is something I like talking to people and also it's something I'm trying to work on.

01:39:54.000 --> 01:40:12.000
I don't know. I agree that I think making the tools more available to communities um is kind of one, it gets people more familiar with machine learning but also like it makes it easier for more people to bring their ideas to the table so

01:40:12.000 --> 01:40:20.000
Great. Well, thank you everybody for being here. Thank you, Cody, Carrie Ann, Amir, and Pharrell for your time and your expertise.

01:40:20.000 --> 01:40:27.000
And so we'll wrap up the session and take a 15 minute break, it looks like.

01:40:27.000 --> 01:40:30.000
Thank you.

01:40:30.000 --> 01:40:44.000
Huge hand. That was amazing. Those were all great talks. That was a great discussion. You guys got the workshop going beautifully. Thank you so much. And thanks to our wonderful moderators as well.

01:40:44.000 --> 01:40:51.000
So as Teresa said, we are on break for 15 minutes until 1130 Pacific.

01:40:51.000 --> 01:41:07.000
For those of you who are speaking in the next session, that would be Nicole, Heidi, Tim, Greg, and Anna's moderator. Feel free to stick around or come back early so that we can check your video and microphone and answer any questions that you have. And then when everybody else comes back at 1130,

01:41:07.000 --> 01:41:18.000
What are you going to remind us to do?