[Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] [Music] foreign [Music] [Music] [Music] [Music] foreign [Music] [Music] [Music] [Music] [Music] [Music] foreign [Music] [Music] [Music] foreign [Music] [Music] hi everyone my name is Rachel silvern and I'm pleased to welcome you to the National academies of Sciences engineering and medicine and to our monthly webinar series climate conversations Pathways to action the national academies provide independent objective advice to inform policy with evidence spark progress and Innovation and confront challenging issues for the benefit of society in keeping with this Mission we're excited to host these conversations about issues relevant to policy action on climate change I'd like to acknowledge that the national Academy's Washington DC headquarters is physically housed on the traditional land of the nakach tank and Piscataway people's past and present we honor with gratitude the land itself and the people who have been at stewards throughout the generations we honor and respect the enduring relationship that exists between these peoples and Nations and this land and acknowledge that the expertise held by different Native communities is crucial to the work of understanding and addressing climate change today we'll talk about the role of artificial intelligence and machine learning in advancing climate Solutions including how these tools are already being applied by the weather and climate science communities as well as future opportunities to develop new approaches to address critical climate questions our conversation will be recorded and available to view here right after the event if you'd like to ask questions please submit them in the box below the video at any time and we'll incorporate them in a dedicated question and answer period in the final half hour we also encourage you to participate in the polls that will appear in that same location and to give your feedback after the event in the survey linked above the video above you'll also see a link to sign up for our Weekly Newsletter about all the upcoming climate activities at the national academies including future climate conversations and next month's climate Crossroads Summit we've also listed a few other academies resources relevant to today's conversation I'm grateful to be joined today by Stephen Lacy a veteran business journalist editor and audio producer who has been covering the energy transition for nearly two decades he's the executive editor and co-founder of PostScript media a media company focused on covering climate Solutions at the intersection of tech business and culture Stephen will introduce our conversationalists and moderate the event thank you again for joining the national class National academies for climate conversations Stephen over to you all right thank you so much for having me so delighted to have this conversation today um really appreciate everyone being here so um as mentioned I don't come at this conversation as an AI expert I come to this conversation with a long background in energy journalism and watching how the energy system and particularly the grid has become so much more Dynamic distributed and digital and as you watch many of these Trends play out in the energy sector inevitably you know you you start asking well how are we going to manage all of this clean distributed energy and that brings you to the the digital layer um underlying the energy system and then it brings you to artificial intelligence and more specifically to machine learning and so that's how I started getting more and more interested in this topic and that's how I initially got introduced to one of our guests here priyadanti who was writing about this and talking about this and a few years ago I um interviewed her as I was exploring this topic more and she she really help me understand the range of applications and you know I have since gotten much deeper into this space so I'm really excited to explore how it's actually being used today and what's interesting to me is is that um suddenly everyone is talking about AI right much of this is just because of the explosion of generative Ai and tools like chat GPT and Bard M it's because it's a tool that we can all experience and it feels like magic when we use these generative AI tools uh mostly for Creative purposes and uh you know coming out of the enthusiasm and crash of the web 3 space all of a sudden you have a bunch of venture capitalists investing in it you have people talking about it you have people experiencing it um um it's been really remarkable to see how suddenly even after 70 years of innovation in this space now is the moment that we're all talking about it um but I do think it's it's different than other previous technology Cycles because there are so many applications today many of which are for positive social economic and environmental good which we will talk about and so I think this is different from other uh you know enthusiasm or hype Cycles in technology where there are a lot of applications right now and of course we're going to get past this period of enthusiasm and maybe we will sort of head into this hopefully we will not head into um a trough of disillusionment right we're on the slope of Enlightenment and and hopefully we can keep going but um I do think it's a really critical moment for us to get it right get it right to explore real applications for companies to integrate this correctly and for us all to Grapple with um what it means um uh so so you know there's so much to AI Beyond generative AI there's reinforcement learning computer vision Predictive Analytics digital Twins and these can all be applied in energy climate and weather applications um we we have uh I'm the co-chair of the transition AI conference and we've been focused on energy use cases and we've identified dozens of use cases in energy their existence today and right now now um it's you know it's a it's a market worth about 13 billion dollars and um you know a lot of the companies that are building Technologies today may not be around but there's a really critical underlying technology piece here that is so important for us all to understand and so we're going to talk today about the research about the history of research about why now is the moment we're focused on it and about where these tools can be most helpful because these are tools and now we have the moment we can seize this moment to use to maximize their benefits um for for the climate for the economy and for people so let me introduce um the the two folks here that you want to hear from again Priya Dante is someone who I've relied on in the energy space to really understand where AI is headed she is an incoming professor at mit's Department of electrical engineering and computer science she's the co-founder and executive director of climate change AI a really fantastic organization that uh non-profit organization that's catalyzing impactful work at the intersection of climate change and machine learning and her research focuses specifically on machine learning for forecasting optimization in in controls in high Renewables power grids hi Priya hi Stephen and uh Amy McGovern and uh to Dr priodonte and Dr Amy McGovern uh is a professor in the school of computer science in the school of meteorology at the University of Oklahoma um and Dr McGovern is also the director of the National Science foundation's AI Institute for research on trustworthy AI in weather climate and Coastal oceanography known as ai2es her research focuses on developing and applying trustworthy Ai and machine learning methods primarily for severe weather phenomena and I came to your work actually Amy through following all the news uh in this space and you have been cited so often and many of much of the reporting on um uh AI for weather forecasting and a lot of the new tools that are emerging so I came to your work through that and I'm really delighted to speak to you today I'm glad to be here so I want to hear how you both got into this space first um Priya why don't we start with you what is your how did you get into this world yeah so I knew that I wanted to work on climate change related topics um kind of from my first day of high school where I had this amazing high school biology teacher who turned the first two weeks of our biology class into a climate and sustainability curriculum and we learned about different topics in this area right water management and air pollution and climate and it really struck me at that time how climate was a huge issue of not just sort of environment in the sense of Flora and Fauna but also of you know personal well-being and of equity given the just distortionate impacts of um that climate will have on you know disproportionately disadvantaged people around the globe and so I knew that I wanted to work on the topic I didn't know how exactly I would do that um when I got to undergrad I fell in love with my computer science classes and was just trying to figure out a way is there a way that I could bring those two interests together and so towards the end of undergrad I ended up discovering this paper called putting the smarts in the smart grid by a group of researchers at the University city of Southampton that talked about ways in which Ai and machine learning would be critical to helping us manage increasingly distributed digital and renewable grids and I got hooked I traveled for a year and interviewed people about power grids as part of a fellowship and started my PhD working on it and we that's how I got here and then so when you formed climate change AI what were the knowledge gaps you were trying to fill why did an organization like that need to exist yeah so I mean maybe get my own personal story there so a couple of years into my PhD I was you know really enjoying and feeling like the work I was doing in Ai and power grids was you know impactful and you know something that I enjoyed doing but it also felt lonely um I was sitting in a computer science department you know uh kind of alongside people who were working on different things and were incentivized by different things and and who were publishing more quickly because they were able to work on Benchmark data sets that already existed versus things that that didn't um and I also was sitting in a um a department called engineering and public policy where there were lots of amazing folks working on climate and energy problems but where they were not necessarily focusing on you know algorithmic development or computational development in the same way I was so I personally felt a little bit of a lack of community in my kind of specific area and was fortunate to stumble upon some other folks at a conference who were coming from the machine learning side and hadn't yet been working on climate but we're really trying to make their way in as well as find some Community among some of my colleagues who were working in the climate and energy space and we're starting to see the large streams of data that were becoming available and that could help fill gaps in policy relevant data and um that they were trying to understand you know where does machine learning play a role so basically there were these gaps I think on a couple of different sides where there were many machine learning folks who wanted to get into this space and didn't know how climate folks who were trying to upskill themselves in the relevant machine learning and wanted to kind of understand how and where to do that and others I think like myself who who wanted to find that sense of community and um these are the kinds of gaps that that climate change AI was started to try to address so Amy how did you come to the world of AI and then give us an overview of the applications that you're working on at ai2es two rather separate questions um let's start with your personal your personal entrance first so I'm actually all of my degrees are in computer science they're all in AI so AI the AI part was first before the other part um I got into ad because I used to like to fly a lot of games and uh I'm a little bit older than Bria and the game's AIS were terrible and I thought I can do better than this I could make something that does better than this and that's what got me interested in AI um and and to that in college you know as soon as I took an AI class this is perfect that you know I want knew I wanted to make smarter computers going into college but didn't really know what that meant even in fact started out as an engineer and discovered computer science unfortunately was able to switch over um the other half of that is that I also wanted to be an astronaut for a very long time and all the way through grad school actually I was convinced I was going to be an astronaut doing AI for earth science in in being an astronaut only in grad school did I discover that NASA would never take me because I've got you know a variety of things but glasses are a good start right so oh well but you know you can't everybody can't be an astronaut we want to do cool things and so getting into working on AI for Earth Sciences was an easy you know application or I really want to make a difference I didn't I one even though games got me into computer science and into AI they don't really I don't want to sound like I'm disparaging games I wanted to make a difference to the real world games can totally do that there are people who are using games to do better education there's lots of ways in which games can do that but it wasn't my interest um and and to me looking at the natural world and trying to find a way to better improve it was a really natural fit and then we moved to Oklahoma um for our job and uh when we got here I was able to work directly with meteorology and if for people who've never lived in Oklahoma you experience the weather here in a way that you don't anywhere else in the United States we get all of it and we get it every day you get a choice you know we get the tornadoes we get the wind we get the hail and and I really to me it's AI is a way that we can we can't stop any of that but we can make a difference and save lives and so I find it really exciting um and the other half your question you asked was about ai2 yes right yeah and I like priya's story about trying to build community because I was listening to that and thinking that it's kind of the opposite right so we were totally trying to build community but completely different coming from completely different ways I have been just like Priya I'm a computer scientist only I started attending a meteorology conferences because that's where people were you know people were doing any AI for meteorology but there were no computer scientists there I was the computer scientist right and so I was trying to build a community of other people from computer science who wanted to work on these interesting weather problems so same goal as Korea just I wasn't attending the computer science conferences um and got involved in AMS uh the AI committee there for a while I was a chair I was a co-chair I was running the conference I'm now editor-in-chief for the brand new journal on AI for their systems um but you know all of that was same goals was Priya I think it's very interesting um that you know we were trying to build that Community up of people who want to do AI for weather and really make a difference and that's how aits came about because I've been doing this for 19 years I coming up in about a week I've lived I've owned a house in Oklahoma for 19 years I've been doing this that long and um you know we we the group of us have been working together we've been working on different smaller grants when the nsfai Institute call came out we looked at it and we said we can do this this can be what we're doing it can be a bigger scale effort the time is right for us to talk about AI for weather and it really is um so we're just ending year three of being an nsfai Institute um and you asked me like what kind of applications we worked on we work on a variety of applications uh we're the amount of money that comes from NSF sounds like it's a huge amount of money but we're scoped to work on those applications right so if you're not an academic you're not used to the fact that there's overhead and all the things you're trying to pay out of all that so we work on convective weather which is the thunderstorms the wind the hail the tornadoes the lightning um we work on winter weather looking at like frozen precipitation like I'll give you an example of freezing rain visibility we work on Coastal oceanography applications there's a variety of those so looking at fog prediction saving sea turtles there's more um sub-seasonal prediction so looking at things on the instead of like the weather time frame which is you know tomorrow Etc um looking more at the three-month time scale um and then tropical Cyclones so trying to improve our understanding and prediction of tropical Cyclones so we have an audience poll here we have the folks who are here watching are across the Spectrum uh I think weighted in Academia but across industry in government and in Academia so a wide spectrum of folks and then primarily people are slightly familiar with AI so uh there are definitely folks who are fairly familiar here but there's a big chunk of listeners and and viewers who are slightly familiar so we can break down some Concepts now and I want to talk free and I have have done a few interviews now and I've asked I usually start with the same question depending on who we're talking in front of and the question I have for you Priya is just identify the different use cases of AI and machine learning in the energy sector obviously your experiences in um is in the power sector um and and I teased a few in my introduction but can you walk us through the different applications uh the specific AI Technologies in use in energy today absolutely so when it comes to Energy Systems often the problems we're trying to address are first how do we build an energy system that is inherently sustainable so you know with lots of low carbon energy with lots of interconnections to balance between different regions of the power grid so that we can better balance a grid that has a lot of Uncertain things going on then we're talking about when you have a an existing grid how do you actually operate it better in a way that allows you to deal with variable renewable energy that allows you to to adapt to climate extremes and so to do that you have to know what's going on and then how to actually optimize your grid based on what's going on um and then in addition to that you're often trying to create kind of you know clean energy technologies that um you know might improve upon the actual things we can put on the grid so better batteries or better solar panels or you know nuclear fusion Technologies things like that and so um Ai and machine learning is actually being used across all of these areas so when it comes to helping us to actually operate live grids machine learning is used to help us forecast things like how much solar power are we going to have how much wind power are we going to have to help us better balance the grid and it's also helped helped in used in areas like State estimations trying to just figure out what are the voltages on my grid like right now how do I infer that based on some amount of sensor data and then use that information now to optimize and control my grid um in power grid optimization and control and energy systems optimization and control where we often see machine learning views is to try to improve the speed and scale at which we're able to run centralized optimization Problems by basically approximating parts of them or um you know helping to warm start these optimization problems or do other things that help us to basically run these problems at faster scales so that we can deal with the fact that for example Renewables are varying from moment to moment on the grid and then we also as we're starting to see more devices on the grid like batteries and electric vehicles and um you know distributed uh solar panels and inverters and all of these just many many devices that you can't control fully in a centralized manner um then we start to see kind of uses of distributed control and reinforcement learning to try to actually to control these devices on the grid in a more autonomous manner in a way that's still kind of satisfies the overall objectives of the grid um now I would say machine learning has been used more more so for kind of understanding what's going on on the grid than core optimization and control in practice and there are you know reasons for that in terms of what kinds of infrastructure exists to actually test out these methods and which parts of the grid are you know more critical in the sense of if you you know have a wrong forecast is that as bad as if you actually take a wrong action on the grid um and so that's the operations landscape and yeah as I mentioned machine learning is also being used to help kind of um speed up planning models and also to accelerate scientific discovery of new technologies by basically learning from past experiments to try to figure out what experiments you should try next so uh Amy over to you uh you know weather prediction is is a very data heavy um science and what is different with the machine learning tools and AI tools that you're using today like what can you do differently that we haven't always been able to do and can you talk specifically about the different kinds of AI in use yeah okay so what can we do differently we're able to process a lot more data we're all able to handle a lot more data um the numerical weather prediction models as they currently exist without any AI or machine learning models put inside them or also they're just very computationally heavy they're just equations that are being done on code that was primarily written like starting in the 1970s onward they're almost all written in Fortran right it's parallelized it's grid-based computation where you're trying to break the atmosphere up into a bunch of grids and then represent the processes that are happening in those different grids and one answer to AI it's not the only answer by the way this is one answer to your question is that AI is getting used inside those grid inside the parameterizations to make it much faster much more efficient so that we can get better numerical other prediction Hybrid models um done very quickly another answer is that people are training straight well I'll give two more answers so another answer is that they're looking at numerical weather prediction models and doing post-processing that's one of the things that was done sort of traditionally right the nwp is pretty good but it has a bias let's use AI to fix the bias um something that wasn't possible until probably the last six months so if you'd asked me this you know a year ago I would have said no we're not quite there yet is actually just doing straight up AI replacing the model like there is no nwp just doing data driven AI forecasting in the last six months I've been about four big papers that have all come out on archives I even haven't even all hit peer review yet they've all just been coming out on archive left and right um where they're talking about purely global weather driven you know data-driven models on a global scale and those are really interesting they're related to you talked to the beginning about chat CPT right they're they're generative data driven models that are they ready to replace nwp yet no but they're pretty interesting um and they're they're probably a sign of what's to come in the future so I just want to remind everybody who's watching that you can submit questions we're going to take time at the end of this conversation so submit those questions through slido in the box below the video and we will be collecting all those as the conversation goes and make sure to get as many in as possible um so I want to get into like policy and planning um but I do want to just ask one more foundational question I think is really important given the makeup of people in this conversation who are listening and that is why now after decades and Decades of technological progress is this the moment that we're suddenly having all these webinars and conversations about artificial intelligence like what happened technologically and then from a public perception perception standpoint that like made this just break out into the open all of a sudden because you all you have both been working on this for a long time you've seen how technological progress has evolved but now suddenly a lot of people are asking you questions about it so what what has happened in the industry uh and in public perception that has shifted things so drastically um Priya first to you actually Amy do you want to start sir sure I mean I think I think my answer on that one is going to be that what is exactly what you said what's happened in history is it AI got good enough that it's getting implemented behind the scenes in everyday products and people have no idea there is so much AI That's implemented in your smartphone in your smart home devices in your car in your airplane like you have no idea how much this is going on in the background even in the USPS mail sorting system right it's just what happened in the last 10 years has there been a revolution in the operationalization of some of this AI you know we're talking about sort of the Forefront of it the new methods that are being developed not all of those are operationalized but the stuff that was being developed many years ago is now working well enough that it's it's AI is just all around you and I think that's why we're hearing so much about AI because in addition to all of that sometimes it goes wrong and I know you're going to talk about that later and we're going to talk about the sort of ways in which it can go wrong in ways in which we need to be responsible and ethical about it but I think that's part of why you're hearing about it now is people are waking up to it because it got quietly implemented and then it started quietly going wrong and people started saying wait a minute when did we give over our autonomy to that right and and now wait a minute why do why is my credit score controlled entirely by an AI you know why is my which controls your rent which controls your ability to get a car which control you know everything of your life is that I think okay I mean I I've been around long enough that I've watched ai go through cycles and AI I think there's there's a fundamental change here in just how much automation Nai is in use right now it's not academic anymore and the other transformation of course is comput computing power because in the older days the older AI winter that gets talked about right there wasn't the computing power to do some of the stuff that we're doing now oh yeah yeah and that's what I was gonna yeah add as well that I think that you know over the last decade we've seen enough Improvement in computing power and um kind of the ability to store and maintain large pots of data that have also enabled us to see a lot of these advancements go forward and that's not just centralized computing power but things like you know your smartphone can actually run you know some amount of algorithms right that's that's not something that existed you know 10 years ago um your smartphone has more power than the space shuttle like that's that's what's amazing it's wild and it's also a change in the last 10 years you're talking about absolutely and um I think like even in the sort of shorter time that I've been kind of working on AI and machine learning there have indeed been these cycles and waves right I think that as we saw you know in the last 10 years this sort of background implementation I think what GPT has seen now is this front-end implementation that is really visceral to people it's it's almost the same way that the internet has been around in some shape or form since the 60s but it's only in the like mid 1990s that really your average person was able to interact with it through a web browser in a way that really awakened imagination for what could be done and made more people able to participate in it and so in some ways it's funny right because some of the conversations are out Ai and energy that are happening today are in some sense the same ones that were happening you know two three four five years ago but I think just kind of you know with a larger portion of the population actually participating because Ai and machine learning have become so much more visceral recently yeah and in use I mean people everybody is able to use it now right it's the school kids can log into chat GPT and we worry about education although I don't think we should be worried but like you can just use so many different tools that now the public realizes our AI absolutely spot on observations um very helpful and also we just have so much more data to train these AI systems on so um okay so let's get into some implications big picture implications um you're both focused on the ethical use of AI so Amy to you first what is responsible and ethical AI in your opinion oh a definition question so so do you want it for everything or do you want it for you know what we've been working on is responsible for weather and climate right yes and and part of the reason I asked that specifics is that I think that it's it's really clear that for the larger scale applications like I was talking about before like something that's deciding your credit score that essentially decides your life that responsible and ethical AI should involve something about being transparent and being able to be able to understand if it's making mistakes and be able to correct them what's less clear I think in the AI for weather and climate Community is how that sort of applies to them they see the things in the news about oh this one is only recognizing white faces because it was trained on only white faces it doesn't see brown or black faces right and they think oh but that doesn't apply to me right my research is in weather and climate and so there's nothing that I can do that I need really need to worry about and so that's why I wanted to clarify your definition my my answer to your question of responsible and ethical AI is that we need to be thinking about all the implications of what we're doing Downstream we need to be making sure that we're addressing the bias that does in fact exist in weather and climate data and that we're not in fact recreating environmental Injustice issues or you mentioned just a second ago about the amount of data that's being available now turns out there's a lot of data in certain parts of the world there's not a lot of data in other parts of the world right so that could be an example where you could create environmental Injustice issues you train it on one part of the world you you say well this model worked fabulous there you know the United States is well instrumented we're going to deploy it in another country it's not well instrumented without ever checking any verification and now you're killing people because you're making predictions that are wrong so that responsible and ethical AI means that you need to you need to be working with the people in the communities where you're affected you make sure you're you're doing addressing the problems that they really need to be addressed and you're working collaboratively with them and that you're you're addressing the bias and not recreating any injustices and then if you do create an injustice you fix it you don't just say oh I just put it out on GitHub it's your problem to fix the source code we're good that's not what happens right we need to be responsible for all of that which includes you know a whole variety of things documenting and there's I could probably go on for a while for that um but you know I I does that help answer your question yeah I think it's worth actually actually just unpacking what the bot what the bias might be in weather and climate data because I'm a little bit more familiar with what it might be in the Energy System but what what what are those inherent biases and how could they be Amplified so we're working on a paper on that right now um actually uh we have a classification system it's building building on this so nist has a responsible AI framework that they're coming out with and they had a classification system for a variety of biases and we're building on it just focus on AI for Earth Sciences so there's a systemic and historical biases so we have four main classifications I won't give you the whole like subcategories they're four main classifications systemic and historical biases which are um the biases that sort of exist in the data themselves like they're the biases that you might not have collected the data right or they could just be you know there could be historical biases in that people in an institution might collect like label certain data certain ways but they don't maybe it doesn't necessarily agree with other people's labels and you might think well how does that apply to whether and weather and the climate I'll give you one short example which is that there's a preference towards labeling um hurricane initiation or tropical Cyclone initiation to Daybreak because people like to be able to look at the visible satellite that's just a historical bias because they wanted to be able to verify with this or with the invisible satellite right and maybe that doesn't matter to you maybe but it really should because if you're talking about rapid intensification it could happen at any time of the day so that's historical and systemic bias there's lots of data that's not even collected that we can talk about very that's very systemic biases right there are lots of sensors that are in more Rich neighbor more affluent neighborhoods than other fluent countries um the second bias is just sort of the the data bias itself like if you have the data so the first one was you did may not have the data the second one is if you have the data there could be biases there's a lot of biases like physical limitations just poor satellites can be there's biases to people report data they tend to report along highways and Roads or you know things like that um with the affluent neighborhoods that we talked about the third set of biases comes from the AI models themselves they can learn they can just flat out create biases they're not without they're just math um and the fourth set of biases which interacts with all of these by the way is the human biases so the humans choose the models the humans choose how we're validating the models the humans are choosing the data the humans are choosing how we're you know sub-selecting the data handling rare data effects that we talk about in database the humans affect all of it right and then the humans themselves are the ones who are applying the AI models they're not being replaced by the AML they're applying the eye model well if they're under time of stress they have their own non-biases so this this whole it's a cycle of of biases that can affect everything we do for our Ai and we need to really be aware of it I have for people who want to find out more about it we have a paper we're working on it should be submitted within the next few weeks I had to talk at AMS um and you can find it online they're archived and they're available online for free so you could you could find that there as well that's great yeah this is absolutely critical so um that that's a wonderful resource for people to explore uh Priya that there's a lot of overlap there with biases in the energy sector why don't you walk through uh how that factors into your focus on um safety critical bias free systems and energy absolutely so yeah I think in addition to kind of you know as Amy mentioned right data biases kind of modeling biases usage biases um and uh yeah a lot lots of things like that one thing that comes up kind of specifically in the um power grid for example is that there's a certain notion of kind of machine learning needing to operate robustly in a safety critical setting so the idea that if you have a machine learning model that takes an action that is kind of you know does not compute on the power grid is not physically feasible or or you know break something then that leads to you know outages and loss of lives and loss of you know you know economic potential and all of those things that come with large-scale you know power grid blackouts and so this idea of robustness is also I think part of this idea of you know how do you evaluate the use of a model um I think the other thing to to sort of think through is that um you know the ways in which society chooses to kind of use AI is not you know value neutral and Amy got at this in in her remarks as well this idea that um for example AI can be used to optimize uh the use of renewable energy on power grids but it's also used to accelerate oil and gas exploration no no increase the competitive oh did you lose me sorry you froze and I thought it was me sorry uh let me know if I should start over at some points but um what I was saying is that this this idea that um you know AI can be used to optimize Renewables but it's also used to accelerate oil and gas exploration in ways that potentially increase the competitive relative competitive advantage of oil and gas against Renewables um it's also the case that Ai and machine learning are often more able to be leveraged by people who have more money and power in society so buy people in certain geographies to solve the problems in certain geographies and it's also the case that sometimes the way we develop Ai and machine learning even for um kind of communities with fewer resources it's it's done in a bit of a um parachute and helicopter in kind of way where the idea is oh great like you know we have the algorithms and you have the problems and we're going to solve them rather than really thinking about how do you really respect and engage the local expertise that does exist and also build up the local kinds of expertise that maybe are not existing in in the same um kind of volume that you would like but really make sure that you're you're not kind of parachuting in for a specific application but really thinking about how do you ensure that you know a larger set of stakeholders around the world inherently has the ability to contribute to the development of these algorithms and the development of these of these use cases in ways that make sense for their local context rather than just being imported from elsewhere so we have a bunch of questions rolling in and I think that this I want to get to as many of them as possible and some of the questions that have come in intersect really nicely with where I saw this conversation going and so to this question about bias how to prevent it how to understand these models there's a question about interpretability like everyone a lot more people are concerned about the black box nature of these tools people are not always clear about how our prediction is being made when that prediction happens how are we getting to certain outcomes that can have very real world consequences um the grid goes down some you know there's voltage fluctuation get weather prediction wrong you Amy you mentioned people die people get hurt what like they're a very real world consequences and we have to understand how is a decision made and it can be difficult to understand that decision so how do we build trust in these models and how significant is the interpretability problem I'm glad that glad to answer it first if you want since that's what we do I mean that's trust is the central focus of what we're doing we're The nsfai Institute for research and trustworthy AI right that is the first part of our name um and and a lot of what we're doing is so we're looking at the central question of why users trust the AI and why users don't trust the AI we are looking at that question for professional end users which are a lot of the ones that you're talking about the decision makers right the weather forecasters the emergency managers the they could be the power grid managers not necessarily the general public although that is certainly of Interest right but the the answer of trust is a little bit different and what we've found is it's very interesting how much they care about looking inside that Black Box model the computer science answer is that everybody wants to just be able to see everything that's happening you know they want to be able to here's the often what you see is a picture of a bird and then you know oh it highlighted the wings and this is why it was important right and and that is important but it isn't all what they need to see a lot of what we're seeing on the trust and it's more of not they want they want a couple things they want a history with the model and it's not so much seeing Al is making its decisions but what it's making like case studies and then understanding where it goes wrong because at least in weather everybody knows that the wobbles are all wrong they're just all wrong in different ways and so you adjust for those different wrongnesses right so the AI model is just another wrong model it just happens to be right in certain good ways and being able to create something that lets you sort of see the case studies of how well it's working and interact with the model like here's the AI model it's giving me a prediction what happens if I change the inputs a little bit right that can help really build a lot of trust it isn't so much that they want necessarily to be able to appear inside the model and say this is precisely the explanation of why it made its decision I'm curious to hear the power grid answer to that too because I think they're probably very related yeah I think they definitely are so I would say that there's a kind of different schools of thought that think about some combination of how much do I trust the data how much do I trust and understand the kind of you know nitty-gritty of the underlying model um and how much do I trust the outputs and I think in different scenarios you're actually looking for something potentially quite different so for example um when forecasting is occurring on the power grid and then so you know we forecast solar power and demand um and then if something goes wrong based on that forecast um often the regulator will want to go back to the system operator and understand you know what went wrong why did something go wrong um and actually a lot of the methods that are meant to audit the forecasting model are actually the regulatory methods are geared towards this era when actually a lot of forecasting and power Goods was rule based and so they were able to kind of say oh well well this rule didn't hold right something about you know it was we we thought it wasn't a holiday but there was actually a really popular TV show on so it kind of was a holiday and they were able to really understand that and I think there's this question that I think legitimately comes up which is should we be using the same regulatory audit approach that that we did when we had rule-based models should we just basically be holding machine learning models to the same standard or should we really think about kind of auditability differently right is it maybe now more about what was not in the data that was represented in in the current scenario is it really about the weights of the model or maybe it is about the weights of the model but I think there's something to think about there whereas when you're thinking about stuff like sub second control on the power grid something where it's you're fundamentally asking something to be very very automated it's not human in the loop at every moment what you instead want is some certifiable guarantee on the output so you maybe don't necessarily care how it got there or even that it got to the best point but you want to make sure that it got to a safe point and so what some of my work looks at for example is if you're able to write down kind of a robustness specifications or physical specifications that the output your control action needs to satisfy then can you actually build a machine learning model whose output is inherently constrained to that kind of safe region of outputs so that even if it did something imperfect internally you know that it's kind of imperfection you're okay with that it's not going to break anything and what and and so how do we think about accountability Amy in weather and climate modeling you mean from the AI model yeah if something goes wrong like so so Priya is talking about something where there's a very clear regulator they have jurisdiction over a certain Regional grid if something goes wrong we know you know they there might be questions about how they interpret what happened but you know there's sort of you you know who is going to be responsible for evaluating it like what how do we think about accountability if something goes wrong and um let's say whether modeling under an AI based system that's an interesting question um because there's a big difference in the our AI models aren't really being autonomously used to generate warnings for example right now or gener you know what she's talking about is a sub second control of the grid and there's clearly not a human over the loop in that because it just can't happen right whereas in what we're doing we're providing models to those people to the humans who are then making the decisions so right now the accountability is still the human right the national weather forecast office that issued the the alert or the warning or whatever they that goes out there's still accountability there if we get to the point where we're automating some of that you know right now the in the in the United States Noah is the only one who can issue warnings for example there are TV stations out there that sometimes claim they do but no one is the only one who can actually issue official warnings if we get to the point where those warnings are issued in an autonomous manner then I suspect there will be some amount of Regulation like she's talking about right because you don't want to be issuing mornings or missing a warning more importantly you didn't issue a tornado warning so a whole bunch of people didn't know about it so they all died um or you know yesterday there was a you know there's right now the accountability is at the more at the local level in the event management level I'll give an example yesterday's news I don't know if you've seen that there was a rock concert that I got hit by a hail storm yesterday in Colorado over they had 90 people injured this is not a typical thing you don't typically you're 90 people injured by a hail storm in the United States right but they were at a One Direction concert and they didn't want to evacuate so you know where's the accountability there right I'm sure that the event management folks were trying to get them to evacuate but then you have to also get the people to listen that last mile is really really critical and something that that the accountability question you're asking about is is is we have to figure out how we cross that line into helping make better decisions and better informed decisions so that people do make better decisions and don't get hit by giant hailstones fortunately nobody died but or at least that I've read on the news it was just people who got injured but still 90 people they got hit by a baseball-sized hail that's wild and uh yeah I think this just idea of yeah how do you kind of audit or regulate these what are fundamentally kind of you know technical plus human systems and like where right that lie is is it tricky and even on power Goods there is subsequent control that does happen but there's also decisions that are made in a human in the loop manner on a kind of more um on a less granular time scale so for example when you're making decisions about how to um dispatch power resources on your power grid like a day ahead um that is based on maybe a forecasting algorithm plus some optimization software plus humans who are often correcting the outputs of that optimization software and so in some sense the kind of accountability and the auditing goes across those different processes and sort of where do you assign um the fault and like what is easier or harder to audit right because it is actually really hard to audit why did somebody make a specific adjustment to the optimization problem they can explain it to you but you know the way you have to audit these things looks a bit different well in people's introspections of their own explanations is not are not always correct you know you know oh this is what I was thinking but but you you're not necessarily some of those decisions are made like this and they're not necessarily thinking at the conscious level about some of those decisions it's hard I mean this gets us you could almost take this same question right to how do we hold ourselves accountable for autonomous driving and things like that right where they the car has mostly control but you're supposed to be helping it's it's it's a hard question and one we don't have good regulation on right now either by the way so we have a question about extremes that I think is really important and there's a question about the applicability applicability of AI and machine learning to um problems where data might be relatively sparse something you know they mention extremely rare meteorological events or extreme let's also add extremes and grid management um uh so how do you factor these events in to training a machine learning model for example do you want to start sure I'm happy to so what's some of my work looks at is is exactly this basically you want models that operate well in your kind of average case so when everything is going well most of the time they operate really well but that also are kind of you know resilient to or robust to your extreme case um and do not you know break your grid if something extremely goes wrong and so what what my work really tries to do is is Bridge the machine learning way of doing things which is learn from data to do the right thing in the average case and the control theory way of doing things which is write down what you think the worst thing is that could happen and explicitly solve for a solver that behaves well even in that that worst case so basically this idea if you need some kind of threat model for what kinds of extremes you you think you want to be robust to um and then be able to write down a model that you're able to reason will actually be robust within that that set of of threats and so what some of my work does it says can you actually um write down that that kind of model for what might happen on your grade in terms of extremes can you kind of formulate that in my case as an optimization problem and can you actually embed that optimization problem within your machine learning model so that the output of your machine learning model is outputting something that is still learning from data to do the right thing in the average case but that explicitly is constrained to to deal well with your extremes and in general this idea of physics integrated machine learning or engineering constrained machine learning or just you know domain knowledge informed machine learning is something that I'm a huge fan of because it allows you to really think about what is the existing knowledge that I already have that is not worth throwing away um and also what are the specifications I need my model to meet um and it also turns out that when you do this kind of thing your model actually tends to be less data hungry because if you're using a pure data-driven model you're trying to expect it to learn all of the rules of physics and all of this stuff that we actually already have a handle on in addition to learning how to perform well and you need a ton of data to enable that to happen and it's not even clear you know how much maybe if you go to the infinite data limit you get a model that's learned your physics but you know we don't actually have infinite data in practice and so um yeah I think this idea of just kind of like physics integrated engineering integrated machine learning helps deal with a lot not deal with but work toward towards a lot of these issues of extremes of distribution shift where your future data scenario might not look like your past data scenario and your physics part of the model can capture that even if your data driven part is not and just trying to merge these kinds of thinking increasing extremes oh well you're you ask a slightly different question at the beginning so the increasing extremes versus extremes and data are slightly different questions right I I was going to answer the first one first what's parsley yeah yeah go for it okay because because you know she's talking about extreme events on the power grid and and and handling those in physics informed way but what I was going to say is that you know pretty much at least for the convective phenomenon that I study a lot of those weather phenomena are just extreme or are outliers or they're super rare which I would cause it call Extreme as well in just from the beginning right like tornadoes I could if I want to make a tornado prediction system that is working really well based on the right you know accuracy score in 3D space all I need to say is it's never had a tornado right because over the 3D volume of the Earth at any given time there's no tornado the problem is when you're wrong you're wrong you're really wrong so you know one answer to the first question you ask about dealing with extremes is we've been doing this all along trying to do a lot of variety of you know hail and tornadoes and turbulence and lightning and all those we do a lot of sampling of the data so we do a lot of sub sampling you you focus in on the areas where you you know you're only trying to make your prediction system in the area where you think that there's about to be whatever that event is um you're only training on the data that's that's working on what whatever that event is that you're looking at and that's one way to handle those outlier cases the other half of the question you ask about sort of the distributional shift is a really interesting one right because depending on what it is we're trying to predict the extremes of so for example I was involved um in not ai3s work I'm also a visiting professor at Google for the last two years and and one of the projects we were doing there was is working on predicting extreme heat and that's something that certainly distributional shifting over time right so if you wanted to train on the last 30 years of data are you going to be able to actually make predictions of heat events that nobody's ever foreseen um and and that's that turns out to be really hard which I think is where I'm going to start to say things like what Priya was saying that physics informed knowledge is really important I think you need to really work with your domain scientist so that your AI model doesn't get some the idea in its mind that this can never happen or alternatively that it doesn't predict something that's completely physically implausible which also gets it back to trust by the way that you know if your AI model does something that's completely physically implausible then your end users are most likely to say oh that thing's wrong it's completely out to lunch and throw it completely out even though it was only wrong in one little tiny Corner scenario we've got so many good questions rolling in so I'll pick a question that is um I think targeted to each of you we'll start first with climate Solutions specifically there's a question on the use of AI tools for recommendations on deploying climate infrastructure renewable energy my quick answer is yes there's lots of really interesting use cases from modeling like rooftop solar um to you know targeting buildings for retrofits and then you sort of run into the normal things that slow those down you know building regulation um interconnection cues things like that but uh yeah there are a lot of really interesting applications here I'd love to hear your take on this Priya yeah maybe I want to just gi
2023-06-28