AI Growth, OpenAI's Smart Neurons, Disease Sniffing AI
34:19 Share

Hello and welcome to Skynet Today's Let's Talk AI podcast, where you can hear from AI researchers about what's actually going on with AI and what's just clickbait headlines. This is our latest Last Week in AI episode in which you get a quick digest of last week's AI news, as well as a bit of discussion between two AI researchers as to what we think about this news.

To start things off, we'll hand it off to Daniel Bashir to summarize what happened in AI last week. And we'll be back in just a few minutes to dive deeper into these stories and give our hot takes. Hello, this is Daniel Bashir here with our weekly news summary. This week, we'll look at a Clearview AI story, two stories from Facebook, and news from the ACM Fact Conference.

First off, the Illinois Biometric Information Privacy Act prohibits the collection of data like images, fingerprints, and iris scans without explicit consent. Clearview AI, whose controversial operations clearly flout these rules, plans to challenge the law and the Supreme Court.

As OneZero reports, it claims that previous interpretations of the act are in need of clarification and lack consistency. Meanwhile, international regulators have found the startup's technology to breach privacy laws and have taken action, including Canada and Germany. This is only a filing and does not guarantee we will get to see a Supreme Court case, but it would be a first. The court has ruled on matters of data privacy, but not specifically on facial recognition.

Next, if Google Glass wasn't enough, Facebook is developing its own smart glasses and is reportedly weighing up the legal implications of building facial recognition technology into them. MacRumors reports that Facebook's chief of AR and VR, Andrew Bosworth, told employees during an internal meeting that the company is evaluating whether a legal framework exists that would allow it to integrate the technology into the new devices.

Bosworth has underlined that no decisions have been made on this front yet, and that efforts to develop AR glasses are still in early stages. Facebook first publicly spoke about the smart glasses project last year, claiming the product would arrive early in 2021. The glasses are expected to rival products from Snapchat and Amazon.

Facial recognition and other AI technologies are built on a foundation of data. For image recognition algorithms to learn to correctly identify different people, humans have to spend vast amounts of time labeling images to feed to those algorithms. As Wired reports, Facebook has built a self-supervised algorithm called Seared that has learned to recognize images with very little help from labels.

The algorithm was fed more than a billion images scraped from Instagram and decided for itself which objects looked similar to one another. For example, it collected images with whiskers, fur, and pointy ears into one pile.

When fed a small number of labeled images, including some labeled cats, it was able to classify those piles. Facebook's Yann LeCun has often said that the conventional approach of teaching algorithms using labeled data won't scale, and long-term progress will depend on algorithms like Seer.

Other researchers have said that SEER boasts impressive results and that the approach will allow us to take on more ambitious visual recognition tasks. But there are limitations. Such algorithms require vast computational power. Finally, VentureBeat reports that the ACM Conference for Fairness, Accountability, and Transparency, or FACT, has suspended its sponsorship relationship with Google.

This decision comes on the heels of Google's firing of Ethical AI leads Margaret Mitchell and Timnit Gebru and its reorganization of the Ethical AI group. Conference Sponsorship Co-Chair Michael Ekstrand stated that having Google as a sponsor would not be in the best interest of the community and impede the conference's strategic plan.

But this doesn't mean that FACT will no longer be supported by Big Tech. DeepMind, another sponsor of the conference which had its own AI ethics controversy this year, is also a Google company. FACT has also sought funding from Microsoft, the Ford Foundation, and the MacArthur Foundation. According to the FACT website, Gebru, who helped found the organization, continues to advise on data and algorithm evaluation and is a program committee chair.

Mitchell is also a conference co-chair and a program committee member. That's all for this week's news roundup. Stay tuned for a more in-depth discussion of recent events. Thanks, Daniel, and welcome back listeners. Now that you've had a summary of last week's news, feel free to stick around for a more laid-back discussion about this news by two AI researchers.

One of them is myself, Sharon, a fourth-year PhD student in the machine learning group working with Andrew Ng. I do research on generative models and applying machine learning to tackling the climate crisis as well as to medicine. And with me is my co-host. Hi there, I'm Andrey Konarkov, a third-year PhD student at the Stanford Vision and Learning Lab. I focus mostly on learning algorithms for robotic manipulation and reinforcement learning in my research.

And as usual, we're going to go ahead and dive straight in into the discussion of last week's news. Starting off with our first major bit of news is the 2021 AI Index, which says that there is major growth despite the pandemic. So, yeah.

The AI index is a super interesting effort at Stanford and elsewhere to document and kind of quantitatively follow the development of AI. And so they collect a whole bunch of metrics, such as a number of papers published, student enrollment, companies, etc.

founded press, a whole bunch. These reports are hundreds of pages. And yeah, the interesting thing here that is noted is that private investment in AI substantially increased despite the COVID crisis impacting the economy in negative ways.

Other interesting bits here is China surpassed the U.S. in significant scholarly work in terms of their more cited peer-reviewed journals. But the U.S. is still kind of ahead in terms of overall AI conference papers over the last decade.

And yeah, there's a whole bunch more in terms of stuff that is in this AI index. So if you find it interesting what's going on with AI, it's definitely good to just go ahead and Google it and you can find this report and actually look at it. It's free and public out there. What are your thoughts about this, Sharon?

I would say that most of the points in the paper are not that surprising to me, but it was fantastic to see it all summarized together. So, for example, diversity in AI is still low, but it is growing for both black and Hispanic groups, as well as for women.

And also that there is brain drain between from academia to larger corporations, both professors as well as PhD students. And so I'm not super surprised at that. Someone who does not plan to go into academia. So I'm

I and also that like corporations dominate in terms of the tooling that us researchers use. And yeah, it does seem to be all consistent with things that have been going on within the community. And these are trends that I would I don't find particularly surprising. And I think that summarize where AI is going right now. Exactly. Yeah. This report

It doesn't have much that's surprising, except maybe the continued growth of private investment in AI, as we noted. But many other things are kind of just showing that things you might observe yourself, but without quantitatively measuring it, you can actually track it and see that these are actually the case and so on.

So yeah, definitely very cool. This yearly AI index is continuing to track what's going on and report on AI and it's always interesting to take a look. So I would recommend doing that.

And on to a more specific line of research. The next article is titled Multimodal Neurons in Artificial Neural Networks from OpenAI is focusing on their CLIP model where they discovered these neurons that were higher level

And could be more interpretable at a higher level of abstraction. And they call them these multimodal neurons. For example, one such multimodal neuron could be the Spider-Man neuron or the popular Halle Berry neuron from way back when. And so basically they've been discovering how or what clip has been discovering in terms of higher level concepts and patterns.

Yeah, they found some pretty interesting things. So a lot of them relate to geography or facial expressions. And it's very much based on the training data. Of course, they also do see, you know, holes in it. So they see that they haven't found a San Francisco neuron, for example, and that they also have found that, you know, there are biases in the clip model where a Middle East neuron, for example, has an association with terrorism.

Yeah, so I think this is pretty exciting. There is, as you noted, a sort of similar thing here, which is in biological neurons, there is a sort of similar thing. There is this famed Helle-Berry neuron here.

which fires for information related to Halliburton. And that can be images, that can be text, it can be, I guess, memories. And yeah, so anything related to this concept. And this is a bit different. Usually in neural networks, we've seen you train neural networks to recognize objects, for instance, and

they wouldn't generalize necessarily to, you know, if they are trained to recognize faces, they probably wouldn't recognize a sketch of a face or they wouldn't. If you if you give it text that says face, it won't find that to be a face.

Whereas here, yeah, the way this works specifically is they've shown that it responds to both real photographs and photographs of sketches that are just hand-drawn and text that is just images of a text. And they also show that previous models, in fact, don't respond in a similar way. So I found this interesting.

Quite intriguing, I think. It does seem like a sign of progress and qualitatively sort of pretty new

outcome for big neural nets. Of course, Clip was already super impressive in terms of showing that by training this matching of text to images, you could then have very accurate classification, but also generalized better than prior practices. But yeah, I think it's very cool.

Right. I think there are there are a few precedents to this work that I would love to highlight because I think there has been work done around this that might not have as big of a splash in terms of blog posts as OpenAI is able to manage with their marketing team.

And that's that. I think David, yeah, David Bowe's work on rewriting a generative model and also in what a GAN can see or not see. And I think in his work, he's able to find, you know, a tree neuron and be able to turn it on and off to put trees or remove trees from space.

from a generative model and a much smaller one than Clip. And he's also able to rewrite rules, which means it's...

If in the data you don't see horses wearing hats, you see only riders wearing hats, but maybe you can have horses also wearing hats. And he enables an interface for doing some really quick optimization on just one layer to just adapt that rule and change the neuron to then fire in that direction, so to speak. So I would say there's a few...

I don't think that's the only work that has kind of looked into that like higher level understanding of where, what our neurons doing within a generative model or within a model more broadly. But it is cool that in this huge model, we see probably far more things, right? Cause that, that model was much smaller and was only trained on, I think churches or just like small, smaller data sets. So yeah,

So, yeah, this is pretty cool that we're able to see this type of information and it might kind of key into what a lot of the, I guess, more anti-deep learning people were hoping for, which is like representation of higher level abstract, meaningful things. And I think we're seeing how the models are learning them implicitly, maybe not everything and maybe not everything perfectly, but it is learning some of these things.

Yeah, exactly. I think as is often the case with open AI, there's no necessarily new ideas here or novel research, but just the results themselves that they get from scaling up and taking a close look is quite intriguing. And they do have some interesting, very novel things.

which is this new type of attack that they call typographic attacks. So specific to this model, they showed that, you know, common and we've seen before that with neural nets, you can make small modifications to an image if you have the weights of a model and you can get pretty silly kind of misclassifications of, you know, pandas as dogs, right?

if you make small perturbations to an image. But with this model, they showed that if you don't even need access to the model's weights, you can just take a piece of paper, write some text on it, and stick it on an object, and it misclassifies it. So there's an image of an apple, and then they get a piece of paper, write iPod on it, and then the classification is iPod.

So that kind of implies that if we keep growing up these models, then we will have even simpler ways to trick them and mess with them in unexpected ways where you don't even need any code. You can just be able to...

do something simple. And this is also building on some previous ideas of like there were patches that people could stick on like stop signs to make it so they're not recognized. And yeah, it's cool to see this also as a result here to kind of remind us that our models have some failure modes still.

Definitely. I saw that circulate Twitter is pretty funny and people, it's almost like, Oh, if the model implicitly can learn, you know, OCR or like a mapping between, you know, visual text and actual text, it, uh, it will be very vulnerable to these types of attacks where you could just slap, uh, an image of some other word onto an Apple and it'll classify it as that other thing. Uh,

Yeah, it's pretty funny. Well, on to a different use of AI and a different kind of line of work that is a little bit maybe dystopian or strange. We've got the article Meet the AI Algorithms that Judge How Beautiful You Are from the Technology Review. And so, yeah,

This article is all about how there's a growing number of startups offering services that can use AI to give feedback on your appearance, basically. So for instance, there's this company, Quove Studio, that...

Started out as a studio that would airbrush images for modeling agencies, but now it is a facial aesthetics consultancy that promises answers to what makes a face attractive. And so...

Yeah, they offer things like advice on beauty products and tips on how enhance images on your computer. But most interestingly for us, there is a facial assessment tool, which is an AI system that looks at images of your face to tell you how beautiful you are or aren't and then tell you what you can do about it. And so they offer this article, which is pretty extensive and has a lot of details and

Uh, if, if you find this interesting or following, uh, the offer tried this service, um, that, uh, yeah, all, all it required was, uh, washing off makeup, uh, taking some closeup photos. And then there was a support that returned recommendations that, you know, are supposedly to address flaws. So, um,

There was an article about smile lines that... Yeah, apparently they need surgical intervention. There was a full report of surgical recommendations written by doctors. And there's different products, serums. Yeah, so...

I guess it's interesting. I definitely wasn't aware that this was so extensive and apparently there's like 20 different companies doing this as well as an open source tool out there. Yeah, pretty, pretty unusual and probably not a good idea in general, but if maybe useful to some extent, if it's done well, so kind of kind of odd to see this, I suppose.

What's your take, Sharon? I might be able to touch on this a bit more since I may or may not be more of a target user. I definitely can see how this could be useful, especially since there is a pretty big group of people who would love other people's opinions. I definitely feel like this is kind of a...

Slippery slope in terms of, oh, like it's just feeding into the whole like Instagram. All I care about is my looks kind of market and all of that happening, which isn't great right now, especially for kids.

But I can see how something like this would be desirable for people who want, you know, recommendations on what to do to look better or whatever. I do think there is a little bit of a WebMD effect going on. And by that, I mean, the recommendations can sometimes be a little bit extreme. And so like WebMD, who...

you know, that always says that you have cancer, even if it's like, I'm sneezing. It's like, you might have cancer. And so I think there's also a bit of like extreme recommendations. And I worry in this case, there will be very commercially driven recommendations that might not be actually good for a person. And I think like, if it's something like, you know, buying a beauty product, it's not that harmful, but if it's something where, you know,

It's like, oh, well, you should get plastic surgery. I feel like, well, that's starting to cross some lines in terms of like, what is this really doing for people and for our culture and stuff like that? And of course, there are cultures where plastic surgery is the norm and is very accepted and everything that's not seen as kind of a bad thing or anything. And so I think it, but it definitely will change.

things. And we like, we should be aware of that, uh, that it might do, it might change our culture in terms of how we view vanity and everything. So. Yeah, for sure. Um, interesting to hear your perspective here, obviously for me, I'm not a target user. I don't, I even less than most males, I would say, uh, no, but skin treatment or anything like that. Um,

But I have read there are some interesting reports in the past that are related that apparently TikTok and Snapchat, which both kind of present different content, especially TikTok has a famous recommendation algorithm to give you a feed of videos that is personalized and kind of made to appeal and result in likes.

And there have been some discussion and some allegations that the algorithm includes beauty scoring as part of the recommendation algorithm where, yeah, basically the content that gets promoted is for supposedly more beautiful people.

And there, I think, yeah, we're definitely seeing a less positive application where it's pretty problematic and unfortunate. Clearly, it's unfair to penalize people for their appearance. It's kind of crass, especially given that, for all we know, these algorithms have biases. So this article notes that there is...

Face++, which has an open source tool. And there was actually research that showed that it consistently ranked darker skinned women as less attractive than white women. And in general, European like features higher than other types of features. So, yeah, as usual, there's caveats and as usual,

AI can be used in positive ways and in very negative ways. And it requires some, you know, careful thinking as to where you want to draw the line and then how you want to have AI be applied. Yeah, absolutely.

So definitely a touchy area and we'll see where it goes. I know that several beauty companies have thought about this or at least helped with, you know, thinking through, oh, can we recommend some kind of product to you based on

Some kind of virtual kind of like just saying no need to go in store now or have anyone look at you in person, like just a virtual thing to something a bit more extreme, which which could be just judgment of beauty. And I can imagine this happening.

And I've seen this happen a lot for a lot of kids. I think it can be extremely negative. And by extreme negativity, I don't mean just like their self-esteem takes a little bit of a hit. I mean, like it does. It does. Yeah.

it does influence, uh, kind of suicidal thoughts and behaviors, especially in girls. So, uh, that that's really damaging. Um, yeah. So hopefully there's also, you know, thoughts on that going into something like this. Yeah, for sure. Uh, definitely, uh,

area where care needs to be taken, but also it can be useful, of course. You know, we all want to try and work on ourselves and maybe it's easier to ask an algorithm to give you tips than other people, for instance. True. I would prefer the model instead of my mom, for example.

But onto our last article, onto a very different application of AI. We've got AI system can sniff out diseases as well as dogs do from Scientific American. And so this is about a study published in February where a big team reported an AI-powered system that is as accurate as trained dogs at correctly identifying cases of prostate cancer from urine samples.

So one of the researchers from MIT, Andreas Mersin, who was also our co-author, is thinking that technology could be integrated into smartphones. So there could be a tiny sensor with AI software that could help with checking for diseases, as this says is possible.

And yeah, at a high level, the way it works is that cells produce chemicals that emanate from different things like skin, blood, urine. And there's this artificial nose, including this nano nose that were developed by this team that can detect those chemicals as well as dogs. And then I guess the cool thing here is that...

There's a complex pattern and we use chemical mixes that is kind of hard to know. And so training AI results in being able to do it just from raw data, as is usually the case.

Yeah, very cool. I don't know, Sharon, have you seen anything like this or what does this make you think? I've been waiting for this. I've talked to folks about...

you know, how things are at human level. And I, I was just saying, what about dog level? Like dogs have great noses. We should get to dog level nose and for an AI. And this is exactly going in that direction. So I'm excited to see how it progresses. Cause it'll be a different benchmark essentially, right? It'll be dogs. And I think that's kind of cool. Yeah.

If we could, you know, for the species that we know that has the best sense of X, we will try to get AI to get to that X. I think it also is in line with there has been some work around...

you know, sound and smell for detecting a disease in general. And I think people have been developing what's known as a olfactory camera. And this is still kind of in the works. But if we can easily quantify smell and encode smell, then it'll make work like this much easier. And this definitely makes me think of, you know, work that are outside of

vision, uh, especially and, and in detecting diseases. And that can also be sound if we could, or using other sensors to pick up different, different, um,

And so sound, I think people have been trying at least, I haven't seen extremely promising results. I've seen people trying really hard at it and maybe we'll get somewhere. And I think we've mentioned it in a previous session, but, you know, using cough sounds for COVID detection. So it, I think there, you know, there are directions that we can push on that are beyond imaging. And it's exciting to, to see, to see this and also to,

essentially think about this as dog level intelligence or nose intelligence, olfactory intelligence, something like that, you know? Yeah, exactly. It's kind of interesting here. The neonats were trained to imitate the dogs. So, you know, they were actually trained dogs from...

the UK organization Medical Detection Dogs. And so there was a Labrador and Wirehaired Vizsla that were trained to identify urine samples. And then the AI was made to imitate them from data and then labeled cases. And they got to the same level, which was not that impressive or, you know, not...

super accurate there was 71 percent positive cases of accuracy and 70 to 76 on negative ones which was similar to the dogs for the ai so yeah i think i agree it's it's very cool to see ai being trained to reach dog level accuracy and this also notes that

If you have more training, you can get dogs to 96% accuracy. So it is pretty exciting. It seems like this definitely would make for easier ways to detect different diseases. And as we've seen with prior applications of AI, as long as it is carefully developed and there's no bias,

It's kind of an unambiguously good application of AI. So it's always exciting to see progress. And with that, thank you so much for listening to this week's episode of Scana 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 scanatoday.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. Be sure to tune in next week.