Reading Brain Waves, AI Goliaths, & more on COVID-19
43:58 Share

Hello and welcome to Skana Today's Let's Talk AI podcast where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. This week we'll look at a smattering of AI stories and news that are not related to COVID-19 unlike some of our previous episodes. And then, you guessed it, we're gonna have more COVID-19 and AI news.

You can find the articles we discussed here today and subscribe to our weekly newsletter with similar ones at scannettoday.com.

I am Andrey Karenkov, a third-year PhD student at the Stanford Vision Learning Lab. I focus mostly on learning algorithms for robotic manipulation. And with me is my co-host. I'm Sharon, a third-year PhD student in the machine learning group working with Andrew Ng. I do work on generative models, improving generalization of neural networks, and applying machine learning to tackling the climate crisis.

All right. Sharon, I'm curious. Are you getting tired of me thinking and talking about COVID-19 all the time?

So I have definitely been settling into the new norm, I will say, though my lab has been pretty active on working on COVID-19 projects. And my advisor, Andrew, and I, in addition to a few folks at the Stanford Medical School, recently put out an article detailing the gap between successful model countries in addressing the pandemic versus ours right now.

So you've sort of had to think about it a little bit while writing the article, I see. Just a little bit, you know. Just enough to write it and that's it. Just every night, but that's fine.

just half a day. Yeah, yeah. I'm lucky enough in that my research has nothing to do with health. It's just having robots move stuff around, not badly. So I'm trying to get back to the research and have it be my primary kind of thing to focus on. But of course, it's kind of hard to escape the reality of where we are in.

It's really an exercise of how well you can avoid Twitter, I believe. That's a big part of it. And I'm not very good at avoiding Twitter, it turns out. But I guess for now, the hope of this episode is we can avoid COVID-19 a little bit and instead talk about some other AI news that are actually pretty interesting and cool and speak to larger trends that have been going on since...

Since before this all started and will continue going on after. So to begin, we have an article titled Scientists Develop AI That Can Turn Brain Activity Into Text. And it covers the work of Joseph G.

Macon and colleagues from UC San Francisco. They published in Nature Neuroscience a paper titled Machine Translation of Cortical Activity to Text with an Encoder-Decoder Framework. So basically it was a saying you can record some brainwaves

The cortical activity is your brainwaves and transfer it into the text. So it's sort of like reading your mind, but as you'll go into, not really at all. And the way on a high level this works is it was developed by tracking neural data from four volunteers who had electrode arrays implanted in their brain

to monitor epileptic seizures while they were speaking. So they repurposed the sensors that you use for detecting seizures to try and collect data to then learn to map the brainwaves from that to what they're saying.

So right now, the system currently works on neural patterns that are detected while someone is speaking aloud. So you have to be already speaking aloud, but it doesn't capture the actual speech coming out of your mouth. Rather, it's the neural patterns that are going on while you're talking. Uh,

Experts say that it could eventually aid communication for patients who are unable to speaker type, such as those with locked-in syndrome, meaning patients who are perhaps in a vegetative state and cannot communicate but are still conscious. So participants in the study were asked to read aloud from 50 set sentences multiple times while neural data was being tracked. And these participants responded

read aloud sentences from one of two different data sets. One is a set of picture descriptions, so 30 sentences and about 125 unique words, just to give you a sense of how bounded this set is. So yeah, then with the data being collected, it was fed into a machine learning algorithm that converted the brain activity data for each spoken sentence into a string of numbers.

And then to make sure the numbers related only to aspects of speech, the system compared sounds predicted from small chunks of urban activity. Yeah. With actual recorded audio. Uh, and then vast string of numbers was fed into a second part of a system, which then converted it into words. So they used.

kind of several building blocks of modern day cutting edge AI to build this whole pipeline to go from brainwaves to actual text. And this string of numbers is essentially part of that encoding in the encoder decoder network. And think of it as an embedding, for example, like in other neural networks. So that part is not novel, but the data and the preparation that is novel here.

And as for some results, the system improved from spitting out complete nonsense to actually learning how that string of numbers corresponded to words, which is really exciting because by probability, if there are only 50 sentences, that would be a 2% chance at random probability that it selects the right sentence.

Yeah, so it's quite interesting and cool that it got working to some extent. And in particular, it's interesting that here training and testing was done on the same set of sentences.

which could mean that the system learned something trivial. So it just learned, given a brainwaves to figure out which of the sentences it trained on, the person spoke and not actually recognize each of the words from the brainwaves. But there's actually some clues that it wasn't doing that. And one of those is the mistakes it made. So it, it,

sometimes made quite interesting mistakes such as when someone said those musicians harmonize marvelously the system decoded it as the spinach was a famous singer so pretty different but you do have this sort of association of musician to singer and then another mistake there was is a roll of wire laying near the wall

which became, will Robin wear a yellow lily? Almost kind of nonsensical, not very clear how it happened, but it does mean that there's not a trivial solution going on. So correct me if I'm wrong, Andre, but essentially the decoded sentences from the model were not actually in the training set. So it's not an information retrieval search problem going on here.

Yes, correct. These ridiculous sentences were not in a training set. The person spoke reasonable things. So while the input was in a training set, the output wasn't. So the fact that the system didn't just output something from a training set every time meant that it was trying to generalize to some extent.

And what's interesting is that the accuracy of the system was able to beat some previous approaches. So while the accuracy varied from person to person, as the system was trained on each person individually, for one participant, only 3% of each sentence on average needed some correcting. So that's actually quite low. The system would get it right most of them, like 97% of a sentence, right?

would get it right. And this was higher than the word error rate of 5% for professional human transcribers. So that's pretty interesting. That compared to the human error rate was actually lower for this one particular participant.

But the team stresses, unlike the latter, the human transcribers, the algorithm only handles a very small number of sentences, whereas humans, we can handle a very large corpus. And transcribers in particular probably could handle a very, very large corpus larger than mine.

Indeed, yes. So to expand on these caveats and conclusions, it seems very cool. And you can imagine already the clickbait saying about how AI can read your mind and so on. But there's a lot to note here. So to start, the system gets much worse when you go outside of the 50 sentences it was trained on. So it works pretty well at figuring out the sentences, the

that it had already seen from brainwaves, but when you try something new, it isn't quite so good. So that means that the system is likely relying on a combination of learning particular sentences, like just remembering sentences, identifying individual words from brain activity, and then also recognizing general patterns from all this stuff.

So the researchers note in the paper that although we should like the decoder to learn and to exploit the regularities of the language, it remains to show how many data would be required to expand from our tiny languages to a more general form of English. So with just 50 sentences per person, that's not enough to learn something general. It's enough to showcase kind of a proof of concept, which is really what this is.

But what is slightly promising about this approach is that it uses less data to train than previous approaches. It uses less than 40 minutes of training data for each participant, a limited collection of sentences, which is both a limitation here but also an advantage.

And this is compared to the millions of hours that are typically needed for similar systems. But far from a huge major breakthrough, far from quote unquote reading our minds, let's recall that the system is still trained on brainwaves when someone is talking out loud. So really, while this is fantastic research, another researcher has noted that

OK, Google could do just as well. And you could just use a speech system like Google or Siri or Alexa here and would do just as well or better. Exactly. So this is a great example of how you can have really cool research, really cool AI results. I think we can agree this is actually pretty cool. I haven't seen anything like this so far. But it also isn't some sort of world-changing event. This is a small step in...

in this direction is going to take a lot more work, many more papers, a lot more data, a lot more engineering to continue making progress to hopefully in the end end up with a useful system to help those that can use this sort of thing. But for now we just see sort of a slight scientific advancement which is how research really makes progress. Yes and this is also suggesting that perhaps

And this is a Guardian article, again, titled Scientists Develop AI That Can Turn Brain Activity Into Text. The article title is a little bit hyped. It's not just regular brain activity when you're not talking. It's when you're actually talking aloud.

allowed. Yeah, and this is a common thing with media coverage of AI in publications such as The Guardian. Usually the contents of an article are quite solid and they do note the caveats and they do get the details right, but the titles do tend to lean a little bit towards the more exciting aspects and hide away the more practical things that make it less exciting. Yeah.

So that's just something to keep in mind whenever you see a cool headline. It's not enough to see the headline. You need to read the contents of the article to really know what's going on.

And in this case, the article is pretty good at laying out the details. But of course, sometimes you have to go all the way to the paper to really see what those details were and what was overhyped or not. Or hopefully you would tune in to us here on this podcast. Hopefully that's that's why we're here laying out all the caveats and things you should know. And speaking of things that require caveats and should not be overhyped, our next article is

It's titled, AI Can't Predict How a Child's Life Will Turn Out Even With a Ton of Data. And it is from the Technology Review. And it's about a new paper, a new study published in the Proceedings of the National Economy of Sciences that basically lays out that even when a ton of researchers all try to predict the life outcomes of children,

many people using a lot of different inputs from data gathered over 15 years, none of the models that were attempted really worked well at all. So the context is there is this assumption that an algorithm that is fed with enough data about a given situation will make more accurate predictions than a human or a more basic statistical analysis. So this study was

was presented by three sociologists at Princeton University who asked hundreds of researchers about 160 different research teams to build predictive models to predict the six life outcomes for children, parents, and households using nearly 13,000 data points on over 4,000 families.

And so none of the researchers actually got even close to a reasonable level of accuracy, regardless of whether they use simple statistics or cutting edge machine learning. Yep. Yeah. Actually, the paper was titled Measuring the Predictability of Light Outcomes with Scientific Mass Collaboration. It had something like 50 co-authors or probably more, and 160 teams were involved in building these predictive models. So yeah,

It really showcased that if you have a lot of scientists, you can attempt something with a lot of different approaches and kind of make a strong conclusion that something is not doable or not as doable as we thought, which is their main claim here is that it was useful to have so many people to attempt this at the same time.

I do have the comment here, though, that tens of thousands of data points is actually quite small, given how varied people's lives are. So the variance of the data that is required for this prediction task. So I'm not very, very surprised that these models were not able to do that. Yeah, it's true. I think the data points were...

Partially, that's showing how varied or how much of it there was. But at the same time, for modern cutting-edge machine learning, as we know, it requires a lot of data to actually generalize beyond the training set. And in this case, it's likely not enough. And speaking of data, we can actually go a little bit into details on what that is.

So we use data from a 15 year old sociology study called the Fragile Families and Child Wellbeing Study, which was led by Sarah McCallaghan, a professor of sociology and public affairs and Princeton, and one of the lead authors of this paper. So this original paper with collected data sought to understand how the lives of children born to unmarried parents might turn out over time, how having unmarried parents affected them.

And the way this data was collected is that families were randomly selected from children born in hospitals in large cities during the year 200. And then data was collected from these families when the children were of age 1, 3, 5, 9, and 15.

Quite an IRB need there for this data. Yeah, this presumably, I must imagine, was quite an effort to have a 15-year-old data collection effort with many families, many children. So even though there's not much data, it is data that is actually quite hard to get. Yes, it's sensitive data. So the professor of sociology at Princeton said,

And her colleagues then designed a challenge to crowdsource these predictions on six outcomes in the final phase of this study that they deemed sociologically very, very important. And so these six outcomes included the children's grade point average at school, their level of quote unquote grit or their self-reported perseverance in school and the overall level of poverty in their household.

And then given these outcomes, the challenge participants were given subsets of the data to train their algorithms and the organizers held back some of the data to use for evaluation to see how well these algorithms could generalize and make new predictions for people who have not been seen in training data. And then over a course of five months, hundreds of researchers, including computer scientists, statisticians, computational sociologists,

tried their best to make a prediction. So they had quite a while to use all this data to develop a good model.

One takeaway that the article states is that the most complicated machine learning techniques also were not much more accurate than far simpler methods. And that explainable algorithms often have close to the same prediction power as black box techniques like deep learning within contexts where an algorithm is assessing risk or choosing where to direct resources.

where the added benefit of the black box techniques is not worth the big costs in interpretability in the explainable algorithms. Yeah, I think that...

That is one of the conclusions, in addition to the conclusion that this was just not doable with this amount of data. And although we do note that with more data, you could get perhaps better results, this is a useful piece of information for policymakers to know that if you want to predict something very complicated like GPA or other outcomes...

Even with this carefully controlled collection of data that was presumably quite difficult, either because of the amount of data collected or because of the models or both, the outcomes were not so good. So policymakers should be very careful to over-emphasistically adopt AI techniques for predicting various outcomes, given how hard and how unsuccessful this attempt was.

Right. But even as unsuccessful as AI was on this task, here is something AI is good at. In our next article, a blog post titled Unstoppable AI Flywheels and the Making of New Goliaths. Yeah, this one is a little more poetic. Very, very poetic. And this is written by Della Perrault, who is the VP of Research at the AI Foundation.

which is focused on detecting deep fakes and artificial content in general, in speech, language, vision. At a high level, Dell Up details that AI creates engines for relentless optimization at all levels.

And he's previously commented on how AI will permeate pretty much everything we know now. Anywhere we are making a decision, like using a specific heuristic or setting the value of a constant, there is a role for learned models. And he's suggesting that at a high level, the examples that we will be delving into that he details in this article are

Indicate that massive firms or companies like Google with lots of money in compute will be able to increase their advantage over others, among other things, and grow to the size of a Goliath in this AI flywheel, this AI positive feedback loop where it improves compute, it improves data center optimization, and therefore, content.

creates these large, monstrous, big tech companies. Yes. So that's a little bit high level. So let's just dive into more specifics. So for instance, what this optimization and all levels pretty much means is that you can optimize any part of your kind of operational process as long as you can get data and feed it to an algorithm.

So one example from a few years ago was how about DeepMind has optimized their data center power consumption by at least 40% using machine learning. So the implication there is that Google and other large companies can continuously build increasingly complex data centers and optimize the data usage so it effectively doesn't cost more for them. Another example is how

Google use machine learning to optimize the placement of massive neural networks, multiple GPUs, and other computational resources, which means that they have even more of an advantage in training large-scale AI models over smaller players. So this is really interesting. I know some of the folks behind the data center are

optimization, power optimization work. And it actually will be quite big for climate-related activities as well, which is what I do some research on. If they can optimize their data centers, they can also use significantly less power and therefore be smaller consumers of energy. And Google definitely has this...

incentive to do so as well as they are also large owners of energy resources themselves, powering their own data centers and their own compute right now. So it definitely makes sense that they're trying to optimize for this. Yeah, it's worth noting to some extent, of course, this is beneficial not for just Google, but for everyone, for them to use less energy. Another example that is cited here is

A paper just released recently called Learned Lifetime Aware Memory Allocator, which basically is about how you can allocate memory in a computer, in a production server better so that you have faster memory accesses. And so intuitively, you might expect the internet to work faster if memory allocation is more optimized.

So the high level point with these various examples is you can use machine learning in

in various parts where we used to just implement things by hand, whether it was data center energy optimization, whether it is memory allocation to computer, we can learn from what is actually happening in the system and optimize to make it better and better than what we could do by hand. And while this can have many benefits in terms of less energy consumption and faster computation, it also has the disadvantage of

It being easier for large companies that already have a lot of infrastructure to do this, and so it gives them even more of an advantage. And speaking of unfair advantages for giants, giant tech companies, our next article or blog post by The Gradient is titled Towards an ImageNet Moment for Speech-to-Text. So speech-to-text has made a lot of progress over the last decade, but it's

It's believed only large companies such as Google, Facebook, and Baidu can provide deployable solutions. And several reasons include the high compute requirements that are usually used in papers erected these artificially high entry barriers. So it's not easy for your lay researchers to really get in online.

on this and improve on these models or even verify these models. Second, speech requires significant data due to the diverse vocabulary, speakers, and compression artifacts. So current existing data sets that are publicly available don't have this diversity as much.

And lastly, there is also this mentality where practical solutions are abandoned in favor of impractical yet state-of-the-art solutions, as the article states. And the authors of this article went on to do a project to try to alleviate some of these concerns.

Yes, so they describe this project in this article. So to be a little more concrete, speech-to-text is when you talk to Siri, when you talk to OK Google, these systems hear what you're saying and then translate it into text and then can respond to you. So all of these systems like Siri... Sorry, I'm not sure how to help with that, but I'm learning more every day. Okay, great. That should totally go into it. Okay.

Also, why is your spirit a dude? I don't know. I like that voice. So before I was interrupted, I was saying there are these systems by Google and Facebook and all of these large companies. And these systems are proprietary. So the data used to train them. The algorithm used to train them.

we don't have out in the open either as open source code and data or as papers for the most part. Although we know some of the parts involved, we don't know the full details, which really hurts the ability of smaller labs, smaller practitioners to be able to get into the space. And so the article or the offers here

outline some of the steps that I've taken to try and make it easier and move the whole field in direction where it's not just these big players that can, uh, do things here. So the three main contributions made are one introducing the diverse 20,000 hour open STT dataset, which is published under the creative commons, non-commercial license. So it could be used by researchers everywhere. Uh,

The second is that it demonstrates that it's possible to achieve competitive results using only two consumer-grade and widely available GPUs. These are the 1080Ti's from NVIDIA. And thirdly, it offers a plethora of design patterns that democratize entry to the speech domain for a wide range of researchers and practitioners who are not necessarily the Googles and Facebooks and Baidus of the world.

Yes. So going back to that first point of the large dataset, the title here is "Towards an ImageNet Moment for Speech-to-Text." And what that refers to is the ImageNet dataset, which is a very large dataset of images that has been very often used in computer vision and made a lot of tasks a lot easier to do with less compute and less data. And so their main contribution to some extent is they note that existing academic datasets

have a lot of drawbacks. They're too clean, they sound like I came from a studio, they are too focused on one domain, like just phone conversations or news, they're mostly in English, and they're just too small to actually do anything impressive. So because of these drawbacks, about six months ago, these authors decided to collect and share an unprecedented spoken corpus for the Russian language. We'll

with a target of 10,000 hours at first, which is larger than any other open dataset of this sort.

So what this dataset essentially is addressing is how academic datasets right now suffer from drawbacks like being too ideal. They're too clean. They were recorded in a studio. There's no background noise. They're too narrow of a domain. They're only focused on a certain set of, let's say, vocabulary or type of sentence or types of voices, types of certain people talking.

And they're mostly in English as well. Yes. So they collected this large data set as a response to that. And they also in this article describe how they went about basically trying out different AI techniques and experimenting a lot.

And ultimately, we're able to get good performance quantitatively that is on par with published research, but that requires less computational hardware, which is another way in which the big players have a bigger advantage.

So the interesting thing with this blog post is it kind of highlights how to some extent AI has been in the process of democratization. Basically, anyone with some knowledge of programming can self-educate about the latest cutting edge AI ideas and implement some code and try out some ideas. On the other hand, some subfields in AI, such as speech to text,

still make it far easier for academic groups with a lot of computational resources and access to data to make progress while practitioners outside of academia or industry cannot really get in. So it's an effort to try and change that and keep us moving towards a place where big labs and industry can

are not the only ones able to make any progress. So in the spirit of democratizing AI, we want to conclude our set of non-COVID-19 related articles with an interview with the Stanford professor Chelsea Finn, entitled, Women in AI, I Certainly Feel Like a Minority.

As part of the Women in AI special project, the Medium publication Synced spoke with Chelsea Finn, an assistant professor in computer science and electrical engineering at Stanford University, whom both Andre and I know. Most of her work focuses on the capability of robots and other agents to develop broadly intelligent behavior through learning and interaction.

Yes. So this article went into a lot of details that I think both of us and really anyone working in AI has been aware of, but that is still worth highlighting and making sure we are trying to make progress on it. And so we were

The primary focus here was that women continue to be severely underrepresented in computer science and AI. If you want some quantitative measures, previous analyses as part of the Women of AI series from Sync, found that only 18% of the offers in the 21 most leading AI conferences are women. And as of 2015, women made up just 18% of CS majors in the U.S.,

So 50% of people are women roughly, and only 18% are really publishing and getting degrees in AI and CS.

And the numbers for, I heard, senior faculty are significantly worse as women are just now getting some more and more attention in terms of improving the pipeline. I think one telling perspective that Chelsea gave in the article was at artificial intelligence and machine learning conferences, for example,

I certainly feel like a minority and I know that there have been many situations where I was the only female, for example, speaking at a workshop or speaking at some events.

But she also says that she loves that there are a number of other female computer science faculty at Stanford. So they have community as they build out the next generation of what hopefully AI faculty and AI researchers look like. Yeah, actually, this is something I think that is quite nice about the Stanford AI Lab is in the past several years we've had

multiple recruits of new female faculty. So Chelsea joined the faculty last year and before that we had Jeanette Bogd, who is a roboticist from Europe. We've had Dorsa Satig from Berkeley. So the number of faculty has expanded pretty rapidly and that is certainly pointing us in a nice direction.

I think Stanford in particular, the computer science department has put pretty strong emphasis on diversity recruiting in general. And I've been very impressed with them. I've chatted with folks at other institutions where that is not the case at all.

And so it's a very different vibe. And so when you intentionally do try to recruit for minorities, for example, that hopefully will create a much better environment where there is that community. Yes. And Chelsea also mentioned in the interview that she has worked with programs such as AI for All, which is meant to help young women get into AI and CS.

And this AI4ALL program was actually started by one of the professors in my lab, Fei-Fei Li, in a bid to really try and change the situation and make it easier for minorities to enter the field. So the more you allow women and minorities to have high-level positions, the easier, hopefully, it is for the situation to change and for the numbers to improve. So it is nice, at least, that...

We can add a positive note that things are changing. And if we all do keep in mind that we should try and improve things, they do improve. And I've definitely contributed to the AI for all efforts in the sense that I also gave a talk recently at a high school that was for an event that was sponsored by AI for all. So they have a pretty wide reach, at least around here.

Yes, so it is a nice thing to note. And us as PhDs and researchers, we should be working towards improving the situation. But now for the moment, you've presumably been waiting for our weekly update on COVID-19 and AI, which we'll try to keep relatively brief. The first thing we have here is an article titled Using AI Responsibly to Fight the Coronavirus Pandemic from TechCrunch.

And essentially, it is talking about how, as we've talked in previous episodes, there's many AI applications being used by various agencies in multiple countries. Things like surveillance, monitoring of a disease, detection are things that everyone is trying to do very rapidly. And so this article is about recommendations from a UN committee about...

how to make sure to remember to be careful about human rights while building LGBTI systems.

Yes. And they particularly note that we must not forget that it's that the use of AI can raise very real and serious human rights concerns that can be damaging and undermine the trust placed in governments by our communities. And they lay out several suggestions on responsible use of

For example, data anonymization. So some countries are tracking individual suspected patients and their contacts, while other countries are collecting anonymized data to study the movement of people in a more general manner.

They also lay out a purpose limitation, which is basically ensuring that personal data is collected and processed to track the spread of the coronavirus, but it should not be used or reused for another purpose. And this is similar and along the lines of GDPR or the EU's General Data Protection Regulation, a policy that has been out for some time now. But it's time for this to become a global principle for AI.

Yes. And some other things they mentioned is that it's a good idea to share data across countries, not kind of hoard it. And ultimately for these systems that are put in place, it would be nice if they are put in place with time limitations so they don't necessarily stick around. We don't have pervasive surveillance coming around.

But on top of discussing these recommendations, the article also notes that the United Nations Interregional Crime and Justice Research Institute has established a specialized center for AI and robotics in The Hague. And that is one of the international actors dedicated to specifically looking at AI in respect to crime prevention and control, criminal justice, and rule of law and security.

So the center is meant to help various national authorities and to understand the opportunities that these technologies present, but at the same time, navigate the potential pitfalls associated with these technologies. So it's not necessarily a center just for COVID-19 and AI. It's more broadly about AI. But in this particular moment, and a lot of AI systems are being deployed very rapidly, the center is more important than ever.

But all that being said, AI does have a useful role to play. And that was detailed just this past week at a conference held by Stanford virtually over Zoom, specifically by Stanford HAI. The conference is titled...

Artificial intelligence and COVID-19, how technology can understand, track, and improve health outcomes. And so this conference convened around 30 AI researchers and experts to meet virtually and discuss ways AI can help understand COVID-19 and potentially mitigate the disease itself.

and developing public health crisis. And so we discussed some of the takeaways from this conference, which we did touch on a little bit last week in terms of this is a very ad hoc conference that they turned around from a completely separate topic just to address the current issue at hand.

Yes. So this conference happened just last Wednesday and it was actually open to the public. So you could just tune in on, I think it was YouTube. I tuned in a little bit myself. And the fun part about the conference and also the article that summarizes it, which we are covering now, is there are a lot of individual researchers and projects highlighted specifically.

So we're just going to go in a rapid pace and go through them because it's interesting to see how much is going on. So for one, infectious diseases data scientist Lucy Lee of the Chan Zuckerberg Biohub says her organization is developing a tool to estimate unreported infections. And then at Stanford, associate professor of medicine Nigam Shah and colleagues are honing in on ways data science can respond both operationally, so in the sense of how many patients

will our region have, how many ICU beds will be needed, and clinically, which means who do we test of the patients. And they also point to critical regions for further research, which is which drugs can help us.

At Carnegie Mellon, Statistics and Machine Learning Associate Professor Ryan Tibshirani's epidemiological forecasting team has shifted from studying the flu to COVID-19 to predict short-term forecasts that will inform public health officials in making policy decisions.

Meanwhile, Tina White, a Stanford Mechanical Engineering PhD candidate, designed an open source app to track the spread of COVID-19 using anonymized Bluetooth data.

HAI co-director Fei-Fei Li's research offers an AI approach to helping senior citizens stay in their homes. Sensors and cameras could send valuable information about sleep or dietary patterns, for instance, to clinicians in a secure and ethical way.

And the conference was primarily about academics, but not only academics. So there was also discussion of how QRI co-founder Xavier Maturian told how his company's machine learning tools create personalized diagnostic assessments.

And then from the company Kaggle, Anthony Goldblum offered some descriptions of how they're offering kind of public data and challenges to let people collaborate on tackling the crisis. There was also work on finding a cure for COVID-19.

Binbin Chen, Stanford genetics MD and PhD student, as well as one of my good friends, says vaccines are among the most powerful ways to curb a pandemic and prevent its recurrence. His team uses AI to examine fragments of COVID-19 protein to determine how they might apply to potential vaccines.

And Stanford bioengineering research engineer Stefano Renzi is examining existing drugs that can be repurposed to combat the disease using natural language processing, protein structure prediction, and biophysics to identify potential drugs. Yes. So all this together just goes to show that there are a lot of applications beyond surveillance. There's a lot of promising uses for AI. Of course, these are being developed. Still, they aren't mature, but...

A lot of researchers, a lot of academics, a lot of industries working on it, and hopefully AI will have a positive role to play. And if you want to read more, you can just Google human-centered AI institutes, Stanford, COVID-19, and find all these details and even watch your full conference, which is recorded for your viewing.

So thank you so much for listening to this week's episode of Skynet Today's Let's Talk AI podcast. You can find the articles we discussed here today and subscribe to our weekly newsletter with similar ones at SkynetToday.com. Subscribe to us wherever you get your podcasts and don't forget to leave us a rating if you like the show. Be sure to tune in next week.