AI Nirvana Song, Facebook Fairness Dataset, No-Code AI
36:14 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 takes.

Hello, this is Daniel Bashir here with our weekly news summary. This week, we'll discuss more news from Google and Clearview AI, emotion recognition, and AI for drug development.

First off, the Gebru saga continues. After the dismissals of Timnit Gebru and her ethical AI co-lead Margaret Mitchell, multiple Google employees have left the company. Now, according to Bloomberg, research manager Sami Bengio, who oversaw Google's AI ethics group, resigned on Tuesday, April 6th. Bengio joined Google in 2007 and managed hundreds of researchers in the Google Brain team.

Bengio, who directly managed Gebru and Mitchell and was considered an ally by the co-leads, announced his departure in an email. Bengio had developed multiple popular machine learning frameworks including Torch and TensorFlow, and was instrumental in helping to build Google Brain. Next, BuzzFeed News has more news for us on the Clearview AI front.

While the NYPD stated in early 2020 that it had no relationship to the controversial facial recognition company, documents from a new public records request reveal that the NYPD used Clearview as early as 2018 and has maintained a relationship with the company.

The documents include emails and official contracts authorizing a Clearview trial from December 2018 to March 2019. The NYPD's relationship with Clearview included in-person meetings and customer support from CEO Juan Tontat.

On Monday, April 5th, an NYPD spokesperson did not answer questions about the interactions and described facial recognition as a "limited investigation tool." During the pandemic, technology companies have pitched emotion recognition software for monitoring workers and even children remotely.

As researcher Kate Crawford writes for Nature, one example is the system Four Little Trees. Developed in Hong Kong, the program claims it can assess children's emotions while they do classwork. A 2019 review of the literature found no reliable evidence that AI can detect emotions, and there is growing scientific concern about the use and misuse of emotion recognition technologies. Prominent researchers already support regulation for such technology,

As Crawford writes, "Just as countries have regulations to enforce scientific rigor in developing medicines, so should they have regulations about technology making claims about our mental states." A while ago, Google DeepMind claimed a solution to the protein folding problem, stating that their AlphaFold algorithm could be used in the future to aid drug discovery.

Recently, more progress has been made on that front. As SciTechDaily reports, thanks to work from the Chalmers Institute of Technology in Sweden, AI can now generate novel, functionally active proteins. One of the researchers states that this work offers the potential for numerous applications, including faster and more cost-efficient development of protein-based drugs.

"Among other benefits of the AI-based acceleration of protein engineering is driving down development costs and realizing environmentally sustainable industrial processes and consumer products." 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 the 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 is myself. I am Dr. Sharon Zhou, a graduating fourth year PhD student in the machine learning group working with Andrew Ng. This means I defended, by the way. I do research on generative models and applying machine learning to medicine and climate. And with me is my co-host,

I am Andrey Karenkov, not a doctor yet, but hopefully on the way there. Hopefully, yeah. I'm a third year PhD student at the Stanford Vision and Learning Lab. I focus mostly on learning algorithms for robotic manipulation and a bit of reinforcement learning.

Alrighty. So, um, as usual, we're going to go ahead and dive straight in to discussing last week's interesting news with respect to AI. Starting off with this first article shedding light on fairness in AI with a new data set. So this is quite an interesting article. I think, uh, that's talking about a new data set released from Facebook.

And basically in their press release, just quoting here, Facebook AI has built an open sourced a new unique data set called casual conversations consisting of 45,186 videos of participants having non-scripted conversations.

It serves as a tool for AI researchers to surface useful signals that may help them evaluate the fairness of their computer vision and audio models across subgroups of age, gender, appearance, skin tone, and ambient lighting. And kind of a cool thing here is that these are paid individuals who explicitly provided their own age and gender information.

And this is apparently something new and then different from what's existed. So that's kind of a fairness angle there. Yeah, I think there's a bunch of interesting details that you saw here. What sort of stood out to you, Sharon? Well, what stood out to me was that

AI researchers are now thinking a bit more intentionally about their data sets and what licensing or using their data set means and what is allowed, what kind of applications downstream use cases are allowed and what aren't. You can't predict certain things from this data based on the user agreement. And I think this is a step towards, maybe the government hasn't figured out regulation yet, but people and researchers are starting to think about this a lot more and starting to set

A really important precedent, I think, in how we release datasets now where we're thinking about, you know, bias and fairness a lot, lot more. Yeah, for sure. This is quite interesting. First, because that's sort of the intention of the dataset, right? To evaluate the fairness of different models across different subgroups. So trying to catch bias in models, which is obviously a problem we've seen a lot of.

And as you said, another part here that's interesting is that there's actually a data user agreement that doesn't allow certain tasks, like training models that identify gender, age, and skin tone. And we've seen quite a few controversies arise, especially because of applications of AI that have

been questionable in terms of should you even do this, even if it sort of works. So you've seen like predicting sexual orientation from your face or we've seen like generating an image of your face from your point of voice. That was actually both of those were papers that I think were published in

and led to some criticism. At good venues too. Yeah, they were published at good venues too. It was very controversial and by people who are known and so very controversial. Yeah, and as you said, this seems like a good step or maybe a sign that as a community, we are sort of taking those controversies and learning from them in some interesting ways. I certainly haven't seen

Right, right. Exactly.

And before we go on, there's a couple other neat things in this press release from Facebook. So they also mentioned that one important application here is deepfake detectors.

And they point out that, you know, they are already different state of our detectors, but we are still open question, including how well these detectors perform across different subgroups like age, gender and skin tone. And yeah, they say here that it's important that this works well for basically all subgroups.

And they actually have a deepfake detection challenge, collaboration with other industry leaders. And I guess the idea here is that Facebook AI with its dataset is aiming to catalyze that in particular. And they further say that this release is part of Facebook's long-term initiative of building AI-powered technologies in a responsible way.

From proactively improving the Facebook experience for communities that are often underserved to inviting new ways to prevent hate speech from spreading. So this is quite a positive and impressive sounding release from Facebook about their fairness initiatives.

We were chatting a bit before, Sharon. We did kind of suspect that this comes as a response to this article that came out a month ago. Correct. So I guess we should mention that there was this quite long article

Let me see what the title is here. From the MIT Technology Review, How Facebook Got Addicted to Spreading Misinformation. And this article was critical, basically, of saying that the fairness subgroup within Facebook isn't doing enough, is overly focused on just bias and not other issues. So, yeah, probably worth reading that as well if you're curious about that.

with fairness efforts and kind of what the AI subgroups in Facebook are doing. Right. Absolutely. I think it's important to delve into what each company is actually doing behind like the name of their groups, because that doesn't necessarily mean that's what they're doing. I guess the most famous one is Google's ethics team. Yeah.

No comment. No more comments. We've talked about it enough, yeah. Yeah, the world has talked about it enough. But probably not enough at the same time. Yeah, you know, it is kind of neat. I think, yeah, this does point to the field of AI ethics and fairness and AI maturing where you have efforts like this that, you know, are actually doing some

We saw a lot of sort of statements of principles and, you know, ethic checklists and things like that in the years prior. And it's neat now to see tools and data sets and techniques being developed more so. And at the same time, there's more reporting that is really high quality about these issues. So overall, you know, both things are interesting.

I think showing that we as a field are kind of progressing in this sense. Right. Absolutely.

So speaking of ethical AI, our next article is titled Government Audit of AI with Ties to White Supremacy Finds No AI. This is from VentureBeat. And so there's a lot to unpack in that title. But basically, news broke this month that the Banjo CEO, Damian Paton, was actually part of a white supremacist group. And Banjo is used for surveillance. And what's more is that...

They are actually not using AI at all behind the scenes. And so a lot of questions around algorithmic bias and discrimination, as well as privacy, have been raised on this company since they both have these ties to this white supremacist group, but also are not really using AI as a technology. And so it's a very suspect. And the state of Utah actually halted their oversight

over $20 million contract with Banjo recently due to this. Any thoughts on this, Andre? Yeah. Surprise or no surprise? A mix of the two, I would say. So we were chatting a bit before we started that, you know, this idea of companies claiming to be AI powered, but in fact, you know, behind the scenes is just...

you know, people running it or some software and no actual AI at all. And the AI is just marketing, uh, might be, uh, more common than is understood because it's kind of hard to know. Um, actually I was talking to a friend yesterday and then found this funny, uh, there's a company that measures your body composition and, uh,

measures your body fat using, you know, fancy sensors or whatever. And their name was, uh, Dexafit and actually tried them out a couple of years ago. And apparently like last year or this year, they changed their name to Dexafit.ai. Oh my God. So there's definitely, I mean, I think maybe we're a bit past this phase, but there was definitely a phase where a lot of companies, you know, putting.ai in their name and, uh,

it was often unclear to what extent this was true. And this example in particular is definitely showing that there needs to be more capabilities on kind of doing auditing when you decide to work with a particular company. Here, like, it's funny, there was a contract, right? A 20 million contract and...

only now you know it turns out there was no ai there at all hey they had great sales people uh fortunately no one on tech uh awkwardly very awkward yep yeah well i'm not super surprised then having no tech part not really surprised since like i don't know theranos is kind of like that but uh

And yeah, but like, there's just like a lot to unpack here that they have a lot of like vaporware, but also like really suspect things. If they don't have the tech, then what what's happening? Is it just like white supremacy group deciding who, like as people, you know, like what's going on and stuff like that. It's, it's like, or who, or what the government or what the police should do, you know, it's really them taking control of that. And that's, that's quite scary to think about as well. Um,

by itself like that. Yeah. It is interesting here. It's also noted that the federal government is currently funding initiative to create tech for public safety. And there is the National Institute of Standards and Technology, which now assesses the quality of facial recognition system and presumably working to create additional standards for things like surveillance, autonomous driving, things like that. So this case, I think, showcases

the need for that and quite dramatically as well. Yeah, definitely. I'm glad NIST is on it. Exactly. Before we move on, something fun here. The audit report that, you know, exposed that there's no algorithm is actually publicly available. It's like 25 pages. You can go and read it. So let me just quote to you kind of a fun bit.

So this is from the Office of the State Auditor by the Attorney General of the State of Utah and is about limited review of Banjo. And we can skip here to there's apparently a system called Lifetime that Banjo was touting.

And apparently Banjo represented the ability for Lifetime to perform live event detection based on integration of data from outside sources such as social media or private security data. Banjo touted the ability of Lifetime to identify child abduction cases, active shooter cases, traffic accidents, event detection and real-time events.

I'm kind of impressed. I know, it sounds pretty impressive. And yeah, these are similar to representations, including in its RFP, which led to a contract. So presumably this is what was the approach to Utah. Yeah, I understand why Utah bought it. I know. And then, yeah, so the response to this from the audit was,

Lifetime, as presented to the commission, appears to be a dashboard of data aggregated from Utah governmental sources such as 911, dispatch sensors, police agencies, and traffic cameras. Banjo, expressly represented to the commission at Banjo, does not use techniques that meet the industry definition of artificial intelligence.

Ouch. That's brutal. Yeah. So for a government report, this is a bit spicy and kind of interesting. For a government report. I think so. I haven't read many government reports. I agree. You're right. Cause like, we don't expect that from a government report. Wow. Okay. Yeah. That is, that is quite spicy. I agree. I agree. Fully agree with that. Um,

Well, onto our next article, also about not coding, but in a different way. So companies are racing to bring AI to the masses with no code software. This is an article in Fortune. Basically, this is not super surprising, but companies are,

are emerging in the market and there's a clear need for basically non-technical or non-AI experts to be able to create their own AI models and systems and be able to run AI, uh, still. So this is companies like obviously.ai, DataRobot, Levity, Clarify, Teachable Machines, Lobe, and Accio, uh, just to name a few. Um, and, and one such, uh,

One such player that was mentioned in the article is Primer, which is a company that is focused on doing NLP solutions, natural language processing solutions, where anyone could kind of run these things.

I don't think this trend is necessarily very surprising, but it is interesting to see more and more people be able to develop their own AI systems. And of course, that comes with issues on its own, right? Because, I mean, the more people kind of in charge of each of these systems or models can embed their own biases in them and take that to a new level of optimization as well.

Yeah, yeah. This is something that's been developing for a while. I think as AI got to be big, this was maybe one of the early sort of types of things industry tried to do is to make it more accessible to regular people to create AI models. And as you kind of explained, there were a lot of these things

companies, you know, at least a dozen, if not more. Yes. Yeah. So this article itself isn't surprising, but it is interesting to read if you're not aware of this. And I think it does point to

Kind of it being interesting, sort of whatever major types of companies and industries. So this is one of them. There's also a lot of companies now on annotating data and collecting data sets. You know, there's probably at least a dozen of those that have popped up.

There is a lot of companies providing APIs for speech recognition, you know, computer vision, kind of these high level standard AI tasks. So, yeah, I've always found it interesting where you get dozens of companies doing sort of the same thing.

which happened with AI very quickly as it got big. And I guess, yeah, it'll be interesting kind of which of these get to be mature and actually turn out to be big as opposed to, I guess, the majority of them won't be around much longer. Right, right.

And it's also interesting to see all these pop up because there are use cases and now it's almost like AI is becoming more of a consumer product as opposed to just like something that Google runs internally. And so it's just becoming more widespread in terms of what people use. Yeah, this makes me think, I think the worst thing that could be the result of this being so accessible, I think with big companies,

They have the money and generally getting expertise, people who know AI. And there's so much hype around AI that there's a lot of people graduating, having taken classes and things like that. And Google and Facebook have internal classes as well. So there's many ways to learn. But what I suspect is with these super accessible things, we're going to get more...

Kind of individuals, small teams, building little web apps and apps that are very dubious. So this reminds me of when there was this genderify where you enter text and guess the gender based on it. That was like really ridiculous. And yeah, I could foresee more of those sorts of things happening.

Yeah, that could definitely happen. You're right. And I think like, yeah, I mean, a lot of these are selling to small businesses as well as consumers. So it'll be a mix of things. And it's just like, well, if everyone has their hands on AI, what does that look like for the world? Yeah, exactly. So yeah, something interesting to look to. It also reminds me,

Personally, I haven't used any of these tools. I mean, you and I share an all too comfortable writing code with AI, I'm sure. Too comfortable is the right term here. We shouldn't be this comfortable, to be honest. For our own good, that is. We've spent too many hours debugging and figuring out different things. It's very sad. It's very sad.

But at the same time, I have used some kind of similar tools. So they are in particular a few tool tools that that make it easy to create AI based art.

And yeah, so I've used some of these tools. I'm blanking on names, but they're quite popular. And, you know, one of them has this idea of evolving art by mixing different things using generative models.

And you just click, you just click on different images and, you know, use some sliders and there's no need to do any coding on your own. Yeah. Art breeder. Yeah. Art breeder. Exactly. Yeah. Yeah. Yeah. So that's the one I've, I've used. And then it seemed really interesting. There's a really thriving community in art breeder. The tool itself has, has really evolved over time. And that's a case where I think no code, um,

AI, which it is in a sense, right? Definitely. Absolutely. Exactly. I don't know. Have you played at all with Artbreeder? Oh, I've played a ton with Artbreeder and I've showed it in my class and in both my classes. And yeah, I know Joel too, the person who made it. It's incredible that

they've developed such a huge community around it. I'm like super impressed. And I feel like, honestly, it's the best way to interact with the generative model. Exactly. Yeah. And reading this just reminds me I should go back and, and play with some more. I've at one point I got really into it and generated like dozens of images. And then, yeah,

got a bit out of it but it'll be interesting like right now it's quite powerful you know you should if you if it sounds cool I would say listeners you should just google art beater and play around with it and be amazed be amazed yes definitely I was pretty amazed even as researchers it's pretty amazing yes I feel like it's just the best way to interact with a GAN because that's what they are

Alrighty, and then onto our last article. Something also kind of fun and also kind of related to art, as we were just discussing. The name of the article is New Nirvana Song Created 27 Years After Kurt Cobain's Death Via AI Software. And this is from billboard.com. And yeah, this is quite interesting article.

So the story here is that there's a new musical project titled Lost Tapes of a 27 Club that utilizes AI to analyze up to 30 songs per selected musician who struggled with mental health issues and died at age 27, which includes people like Kurt Cobain, Jimi Hendrix, Jim Morrison, and Amy Winehouse.

And this is done by Over the Bridge and Toronto-based organization aiming to raise awareness about mental health within the music community that created this Lost Tapes of the 27 Club. And they've already released several songs, including this one that sounds like Nirvana and one other that sounds like it's by Amy Winehouse.

So I'm going to go ahead and cut in a bit of that so you can get a taste of the song and then kind of hear it and see how it relates to Nirvana. And then we'll be right back to talk about it. Got my hands around the gods of hell

All right, you're back. So that I hope you were impressed. I was impressed when I listened to that. It sounded like Nirvana to me and also like a good song in general. I know both of us listened to these before starting. What was your impression, Sharon?

A huge wow. It's awesome. I know it's not all like end-to-end machine, but it is getting there and it's really good. I'm really impressed with it. It does sound like Amy Winehouse. I don't know the other person well, but it does sound like Amy Winehouse to me. Yeah, this is, I mean, we've seen this sort of things before in terms of generating music and even emulating music.

by particular people. So OpenAI had this big project. But what's interesting here is all these prior things, you could hear that they were AI-generated. There was kind of AI artifacts or something like that where even if the overall style and genre and voice were sort of on, they were still quite dirty or...

had all these like weird AI artifacts. Whereas here it really sounds clean. It sounds like a song that could be released. And I would imagine this is partially because they had an actual audio engineer. These aren't just raw outputs of an AI model. They seems like had a very, um,

they figured out how to do this in a good way where they use different AI components. So they used Magenta to get analysis of the compositions of these artists' songs. Then they used a genetic algorithm to create, or they used also a neural network to analyze artists' lyrics. Yeah, so it's another case where the combination of

AI and a talented human who has a lot of skills can be better than just trying to use AI. Yeah. And anyway, quite an inspired project, quite an interesting use of AI. And personally, I think it's an interesting way to raise awareness about mental health for musicians. It's not something I would have thought of. Yeah. I thought it was great that they kind of led with that. I, I,

I really appreciate that. I sometimes like, so I can't tell how like genuine that is, but it is really nice that they do talk about that and like talk through, you know, like this is, this is a huge motivator and maybe a reason why, like in this case, it's,

more okay or at least like more interesting it's to raise this awareness ultimately as opposed to just be like hey check this out and maybe they didn't consent but whatever kind of thing um yeah to me it seems fairly authentic right uh it's actually shocking to think that um

you know, they're working on all these musicians who died at 27. So that's super young. I'm now 28. Right. So that seems like a super early way to die. And it's quite tragic, obviously. Yeah. So, yeah, it's a, I haven't seen, you know, an application of AI that was this interesting and kind of had this sort of

um really mental health oriented goal in a bit and um yeah quite interesting and i would recommend listeners to check it out and you know um listen to our songs and maybe share with your friends because i think it's it's quite cool and um you know i guess it's a good message to spread as well

Yeah, definitely. And I'm impressed Magenta has gotten this far. Google's Magenta team has gotten it this far. I think it's a really cool application. Yeah, I think it also points to, in general, this has been kind of a slow process, but AI for art has been developing, you know, the past decade. So there's many, a fair number of artists that utilize AI,

There's actually a whole website that just has a curated list of all of these applications. And this is showing now, you know, people outside software engineering, people who are not, you know,

software engineers not really talented in that sense who are maybe more musicians are now able to start using tools that are out there to do interesting things which really only started happening you know in the past year or two but I think this is actually quite exciting and maybe not discussed enough in AI what happens when you know it's accessible enough and the combination of human and AI really results in

pretty great things. And yeah, like these songs, again, you might want to check them out with full versions because they're just fun to listen to more than anything. Yeah, they're pretty intense. And so with that cool article, 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 and a review if you like the show. Be sure to tune in next week. Be sure to tune in next week.