Satellite Image Data, Moderating AI Dungeon, Consolidating Autonomous Vehicles
24:25 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 AI researchers as to what we think about this news. Just

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 takes. Hello, this is Daniel Bashir here with our weekly news summary. This week, we'll discuss the White House's AI website, Sammy Benjia's new job, and two autonomous driving stories. To make AI research more accessible across the nation, the White House has launched the website AI.gov.

According to Axios, the White House aims to drum up excitement for AI and broaden educational opportunities. The site's target audience is the general public, and it seeks to make public information on AI more visible to people interested in science. Users can learn how AI is being used across the nation, such as in responding to the pandemic.

The site is also meant to be a tool to advance research by providing researchers computing and data, and to promote equity in the science space by providing information on scholarship and fellowship programs. Elsewhere in the AI world, Sami Bengio, the distinguished Google scientist who managed Timnit Gebru and left in the tumult following her departure, has recently joined Apple.

As Reuters reports, Bengio is expected to lead a new AI research unit at Apple under their Senior VP of Machine Learning and AI Strategy, John Gianandrea. Bengio's new role comes after about 14 years at Google, where he was an early leader of the Google Brain Research team and played an instrumental part in advancing deep learning. Next up, unless you're Tesla, if you're building a self-driving vehicle, then you're probably using a LiDAR sensor.

Many automated driving systems today use LiDAR systems that come from Silicon Valley-based Velodyne LiDAR. But according to Forbes, Pittsburgh-based Argo AI is using its own internally developed sensor to meet its performance requirements. While the highest performance Velodyne rotating sensor was considered insufficient for highway speeds, Argo's LiDAR claims significantly better performance.

Moving over to China, Chinese tech giant Baidu has recently rolled out a paid driverless taxi service and become the first company to commercialize autonomous driving in China.

According to TechExplorer, this is the first time Baidu has had a demonstration with no safety driver sitting behind the wheel. Instead, the safety member sat in the front passenger seat. Up to 10 robot taxis are operating simultaneously in an area of 1.2 square miles, transporting passengers between 8 stops in Shougang Park in western Beijing.

Baidu is known for its search engines, but began testing autonomous driving on the open road last year. Its robo-taxi service has carried more than 210,000 passengers in three cities across China and aims to expand to 30 cities in the next three years. 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. I'm Dr. Sharon Zhou, a graduating fourth-year PhD student in the machine learning group working with Andrew Ng. I do research on generative models and applying machine learning to medicine and climate change. And with me is my co-host...

And I am Andrey Kronenkov, a third year PhD student at the Stanford Vision and Learning Lab. I focus mostly on learning algorithms for robotic manipulation and reinforcement learning. And let us just dive straight in as we usually do. First story from the South China Morning Post titled China makes world's largest satellite image database to train AI better.

And this is all about the FAIR 1M database, the fine-grained object recognition in the high-resolution remote sensing imagery database released from the Aerospice Information Research Institute in Beijing.

And yeah, as it indicated, this is kind of a way bigger satellite imagery database. It has 15,000 high definition satellite images with something like 1 million scenes. So it has very, very precise labels, for instance, specific models of airplanes, specific types of sports courts, baseball field, basketball court, different

different kinds of roads, etc. So perhaps unsurprising, but interesting in that as with any other field of AI, satellite imagery has been getting more and more explored and we have more and more data sets. And this is the latest example of just getting bigger and bigger. So yeah, Sharon, any thoughts?

So it is a large label database. So it's very densely labeled with one million labels. I would say the number of scenes, 15,000, is actually not that large. Probably what's remarkable is just how densely labeled things are to help with training these models. And probably what's useful is that these are high definition models.

or high resolution, excuse me, high resolution images to train our models so that they can see things pretty closely. And there are a ton of different objects that you can imagine, as Andre pointed out, you know, it's not just an airplane or road. It's like the type of road, like an intersection or roundabout, the type of airplane, you know, Boeing 747 kind of thing that'll help

uh, these algorithms significantly. One, one thing that I thought was interesting is that they've also started to use, uh, some of this AI trained on satellite imagery China has, uh, to track the speed of city expansion, uh, in Xinjiang, as well as, uh, how movements of wild animals on the Tibetan plateau, uh, have gone, uh, to the construction of different infrastructure, um,

for the Belt and Roads initiative, uh, that China has been pushing out, uh, for, for quite some time. And I mean, what's not said here is like, obviously there are some interesting civilian related issues, uh, uh, with a lot of, you know, the minority groups in China, obviously the Uyghurs and stuff like that, that I imagine might be also used for surveillance here. Um, but

But overall, I think it is exciting to start leveraging some of this giant database and start to put labels on them such that we can actually build our vision models to be much more powerful. Yeah, exactly. I think there are, you know, ways to take this cynically and probably not without reason. But at the same time, this is a data set that will be released publicly.

Actually, this article notes that in May, AI researchers from many countries will compete in Beijing for a trophy awarded for satellite image recognition technology using this database. So the intent here does seem to be generally to drive progress in the field. And there's a lot of useful applications of satellite imagery beyond tracking for satellites.

for, I don't know. I actually don't know the details, but I'm sure they are. So, yeah, it's neat to note, I guess, that

In addition, you know, there's now many, many different types of data sets. And this is one of them that, you know, might be not covered in the news as much, but it's still interesting. Yes. And very relevant to, you know, climate applications and stuff like that. Like I can imagine being useful for some of those things.

But back to more of a classic topic that we seem to get back to a lot on this podcast, enough that we might be getting a bit tired of it, but it is what it is. We have our next article, How to Stop AI from Recognizing Your Face in Selfies from the Technology of You. And this is a pretty short piece that is really just highlighting two papers from Eichler.

having to do with artificial with face recognition in particular how to prevent face recognition from working on particular photos so one of them is a system called fox that uses somewhat relatively established ideas but is more of a system and has a thing you can download on this project website and then another paper is uh

Working with a slightly different, kind of more novel idea, it looks like that might work better at some context. So, yeah, interesting to have an article written about fairly brand new research in this way. What did you think about these two papers, Sharon? These papers were interesting. I think the second one almost fooled me in terms of what it was about because...

Originally, when I took a glance at it, I thought, you know, this is just adversarial examples on your image. But it's not just that. They're actually looking at reducing the error on your image so that...

instead of adversarial examples where we add different noise, you know, to fool the model, to make it think that it's something, uh, for this, it's that it kind of fools the model into thinking it is actually a good example and that there's nothing to learn from it anymore. And I thought that was a clever approach. Um, and the other paper, uh,

does largely fall back on the original type of adversarial examples that we have seen before, but it builds out this entire system around it and then starts to evaluate it against, you know, real data and, and real attacks and see how well it does, does actually work. Yeah, exactly. And, uh,

I think what's especially neat, aside from the papers themselves, is this article notes that this Fox tool has already been downloaded nearly half a million times. And in their paper, actually, just looking at the...

introduction and their contributions, it says experiments show 100% success against state-of-the-art facial recognition services from Microsoft, Amazon, and Face++. And they share their cloaked photos as training data and then apply resulting models to uncloaked test images of the same person. So, yeah, I think it's cool to see maybe

a more, more applied, uh, paper of this sort, less theoretical. Of course, we've seen kind of, uh, similar things, but you know, this Fox paper seems to be much more having to do with real world systems. And in fact, you know, looking at the paper that was originally published in June, it looks like it was also part of a security symposium besides Eichler. So, um,

Yeah, I think always interesting to see kind of things that might prevent facial recognition from being ubiquitous and unavoidable. Still probably, you know, not warranting to not be worried about it, but at least...

It does show that there is some potential in the future to guard against it. Right. I think it's also cool that this is an example of a real world system that takes research one step further into the real world, but still is packaged in the same way as research in terms of a paper. And I hope to see more of this moving forward as we... There's a lot of cherry picked examples when it comes to research, but it

It makes us beg for more. It makes us beg for, you know, does this actually work? And so this, I think the Fox paper really emphasized that.

Well, speaking of, you know, dark things of facial recognition, our next article from Wired is titled It Began as an AI-Fueled Dungeon Game. It Got Much Darker. And this article is about AI Dungeon, which is an online game where you get to basically use GPT-3 or some kind of text generation technology

to make your own adventure game. And you get to just talk with it via text and design your own game. We've covered this before, but recently there's been a lot of talk about how AI Dungeon, because it's just generating different storylines, can generate some pretty inappropriate storylines. For example, I think the biggest one that kind of made the news is

involving, you know, pedophilic scenarios. And that's not great. And it's kind of like, oh, shoot, how much we need to probably bring in moderation, you know, to to help with help with controlling these things. Yeah, this, I think, was a slightly controversial topic. Well, so this article covers that, you know, moderation tools were added by the company that's created the game.

interprompting of OpenAI, it seems like. The CEO of OpenAI, Sam Altman, says that in this case, the moderation decision is not difficult. They basically said you have to do something to avoid this sort of scenario. And the users of this tool actually were pretty unhappy about it, right? In the sense of

Um, you sort of limiting the creativity or freedom of the user base. Um, and yeah, it's on the one hand, it's obvious that you shouldn't allow, um, really inappropriate things like, uh, sexual encounters involving children. I guess what people might take issue with is if these are kind of crude attempts at moderation that, um,

end up overreaching. So for instance, some were cited as complaining that it was oversensitive and I could not refer to an eight-year-old laptop without triggering a warning message. And I think that is really getting into what is tricky here is how do you

really get at, you know, actually inappropriate things versus maybe unintentionally inappropriate sounding things. Where do you draw the line, et cetera? You know, I think there are a lot of questions here and I think it's good that, you know, this is a case study because, you know, at least this is a game, you know, it's, it's this text-based narrative game, but,

As we get ever more powerful systems like Clip, where you can use some, you know, text to describe, to generate anything you want. I could imagine, you know, many more examples where AI can be used to create very inappropriate, you know, yeah, really, really bad stuff. And we're going to need to figure out how to deal with that. And this is kind of a good example of that.

Yeah, definitely. Definitely. Basically, it was it was coming eventually. But yeah, I think it definitely is presenting itself as an opportunity to to, you know, add some reins to the model. Yeah. One of the thing I saw was this argument that people are always going to misuse tools. Right. So, you know, it's up to the user to not do bad things, not up to a tool.

And I did see a good counter argument to that, that, you know, with AI, it may not be the user's intention. Like, the AI could just, like, randomly throw in some sexual stuff, right? And that not being the intent. And also, I thought that was an interesting point and that, you know,

we do need to figure out how to steer our AI systems in some ways, which is, I think, a pretty big problem with AI overall is like, how do you control, know the limits of, validate the performance of your AI systems and

I guess in this sort of narrative context, art context, that takes this kind of meaning. But of course, there are a lot of other contexts like self-driving cars, like reinforcement learning, where there is kind of the same problem of like, you never really know what your AI is going to do. There are some things you don't want it to do, but that's super challenging to actually accomplish.

Right. It's really hard to regulate and give it boundaries when you can't really predict what it'll do or control necessarily. Because I think I imagine, you know, for, you know, explosives or guns or something, we have regulation and we don't just give it to the user and let them do whatever they want with it. You know, there's still some some control over that. Like you don't like hurt another person with it. That's not legal. And so, yeah.

I feel like there are still ways to control it that needs to happen and can happen. But I think it is harder because it's not as clear cut or not as easy to draw the line anywhere for the AI systems. Exactly. And just one more thing this brings to mind is I did see a study, a paper that was released fairly recently that demonstrated that GPG-free has...

a pretty potent anti-Muslim bias. Like it wouldn't take much sort of insinuation or just bringing up random topics could really lead it to spew some fairly negative portrayals, you know, stereotypes and those sorts of things. Like even just bringing up a mosque without a complete can go into very negative portrayals. So...

Exactly. Right. So, yeah, many people have been pointing out there's biases, there's issues of these large language models that will need to be really addressed before making use of them. And this is just yet another example that is already out there, I suppose, and an interesting case where something needs to be done already. Right. Absolutely.

Well, speaking of reigning things in, our last article is titled The Autonomous Vehicle World is Shrinking, It's Overdue. And this is by The Verge. And the article basically speaks to how the self-driving car world is becoming smaller. There are fewer players. People are merging together. And also, valuations are going down. And so the CEO of Waymo actually was stepping down. That was announced last month. And

The evaluation of Waymo has gone down nearly 85 percent decrease since 2018, which is huge decrease, though it is still 30 billion dollar valuation. And so I think people are starting to get maybe grow more realistic, less hyped about the field. And it is consolidating significantly.

Yeah, it's this is a fun article, I think, mostly because what we discussed last week of Lyft selling off their autonomous vehicle division happened and this this came out right after, possibly in part because of that.

And it notes other things like that Aurora merged with Uber's autonomous vehicle unit, delivery robot startup Nuro acquired self-driving truck outfit Ike. And in general, there's been a ton of mergers, joint ventures and tie ups that, yeah, as it says, there's been a lot of consolidation there.

And I don't think we've seen, yeah, we haven't seen new startups really. Whereas like in 2015, 2016, it seemed like every week there was a new company being founded by some AI graduate students, you know, saying they want to do self-driving. But yeah, it seems there's been a real trend of,

many companies being bought up emerging and now there's really only several really huge companies that are aiming at sort of a big problem of robotoxy self-driving and these are like cruz remo tesla maybe baidu um

But yeah, I think other independent companies are more focused on smaller scale problems that are more sort of logistics or something less challenging than a full problem, which I think a lot of people, you know, the engineers on the ground working on these systems probably are not too surprised. Right. I think.

A lot of people scoffed when Elon Musk predicted that we'd have self-driving cars by 2020 or whatever. And yeah, I guess most people are right. We don't have it yet. And it seems like it's not going to come here anytime soon. But at the same time, we're still making progress and we can still look forward to it. Right. Definitely.

And with that, 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 skynetoday.com. Subscribe to us wherever you get your podcasts. And don't forget to leave us a rating and review if you like the show. Be sure to tune in next week.