Tina Eliassi-Rad

speaker

154 appearances 1 recordings 1 series first heard Jan 2025 last heard Jan 2025

Tina Eliassi-Rad’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas · 301 | Tina Eliassi-Rad on Al, Networks, and Epistemic Instability · 13 Jan 2025 podcast

I'm not going to be a very popular person, but that if you get taxpayer dollars from in your reports to the government, you have to have a section on assumptions and technical limitations. Because the problem is the way the peer review culture goes is that if I have a technical limitation section in my paper, the reviewer will just copy and paste it and say reject, right?

But the federal government isn't going to do that, right? NSF isn't going to do that. NSF has already given you the money and you're doing the annual report. And so it has to be, come on, just be honest, right? Like I did not test this method on biological networks and they're very different than social networks. So like caution,

Yeah, I love that problem. I've thought about that problem a lot. So the issue there is similarity is an eye of the beholder, right? And it depends on the task itself. So similarity is an ill-defined problem. And so you can say, okay, well, I can go with something like an edit distance. Like, okay, how many new nodes do I have to add to graph number two?

And how many new edges do I have to add or remove to make it look like the other graph? And then try to solve the computationally hard problem of isomorphism. In fact, alignment, right? And in many cases, you don't need alignment, right? So, for example, you can think about two networks and you have started a process of information diffusion on it, like you started a rumor, let's say, right?

And you would just measure, like, how similar does this rumor, the same rumor, travel through network one versus network two? And if like, you know, it travels similarly, let's say, you know, I'm going to throw some jargon, like the stationary distribution of a random walker that is spreading this rumor becomes the same at the end. You would say the networks are similar enough. Right.

And so you don't need to have like the sizes exactly be the same. So it could be, for example, you have a social network of France and a social network of Luxembourg and you start a rumor in France and in Luxembourg. And they are processing the same way. And you would say the networks are similar, even though one is much, much bigger than the other.

Yeah, yeah, now the problem with grouping nodes, this is a very important problem and it's been studied by lots of people. Within graphs, it's called community detection. Basically you want to group similar nodes together. Now you can have different functions that you define about what similarity there means. It could mean that these people just talk to each other more, right?

So there's more connections between them than what you would expect in a random world, right? or just more connections between them than other folks. Now, this kind of community detection, Aaron Closet, who's a professor at Colorado, showed that there's no free lunch theorem there. And actually, it was Aaron Closet and others. And I think actually Aaron was the last author.

So I think the first author is Leto Peel. But you know how it is. You usually just name your friend.

My apologies to the other authors. But they showed it in no free lunch theorem, which basically means that it is not the case that there is like one particular group of or one particular collection of nodes that you're grouping that would give you the best or the best. true communities. You see what I mean?

So because when you are doing these grouping of nodes, you have some objective function that you're trying to maximize. And basically the idea is that there is no one peak there. So there's not like one particular community that you can put Tina on and say, okay, Tina belongs here. That's where she has to sit. And so some of that becomes an issue.

But this notion of what does it mean for one network to be similar to another network has its tentacles to community detection, to clustering of nodes, and all of those are ill-defined. So it really is driven by the task at hand.

Yeah, and that becomes what we call the small world problem, right? Or the Kevin Bacon or the Erdős number, right? You don't have to go that far out. to be connected to famous people.

I mean, for downstream tasks that you can like have some, let's say, confusion matrix where you can draw like true positives, false positives, true negatives, false negatives. We're actually very good at it. But if it's about like, OK, I found these communities and do these communities make sense?

It kind of breaks down into whether they're like hard clustering where you put Tina into just one community or you put Tina into multiple communities. And then there's a little bit of just like eyeballing it in a way. If you do not have this downstream task that you can say, okay, here are the true positives, here are the false positives, and so on and so forth.

But in many cases, it's difficult to place a person in a social network only in one community because people are multifaceted.

I think they're in part, they just want your attention. And so the objective function is such that, you know, they just want to hold your attention. And so they will show you whatever necessary that will keep your attention.

And so if they believe that like my tie to Brandon is very strong, that we have a strong relationship and Brandon found these things interesting, then they will show it to me as well to just test it, to see whether, you know, they can capture my attention. And then through that, they can show me more ads.

Exactly. Exactly. And so they kind of go hand in hand. And in fact, this touches on this issue that we have written a couple of times about. There was a Nature Perspective piece a while back and more recently an AI Journal piece on this. in a way like human AI co-evolution.

So if you think about it, when you're using Amazon, when you're using YouTube, when you're using Google, you're providing data for them. We talked about this, right? And they take that data into account and they make recommendations. Those recommendations then affect what you do in the real life. And then you go back and you provide them more training data.

Showing 41–60 of 154 · page 3 of 8