Charan Ranganath: belief

Transcript context

…I don’t know what… Because we’re ignoring all the other cars, but for some reason, the asshole, like a glowing obvious symbol is just right there, even in the periphery vision because again, we’re usually when we’re driving just looking forward, but we’re using the periphery vision to figure stuff out. It’s a little puzzle that we’re usually only allocating a small amount of our attention to, at least cognitive attention to. It’s fascinating, but I think AI just has a fundamentally different suite of sensors in terms of the bandwidth of data that’s coming in that allows you to form the representation that perform inference on using the representation you form. For the case of driving, I think it could be quite effective. One of the things that’s currently missing, even though OpenAI just recently announced adding memory, and I did want to ask you how important it is, how difficult is it to add some of the memory mechanisms that you’ve seen in humans to AI systems? I would say superficially not that hard, but then in a deeper level, very, very hard because we don’t understand episodic memory. One of the ideas I talk about in the book, because one of the oldest dilemmas in computational neurosciences, what Steve Grossberg called the Stability Plasticity Dilemma, when do you say something is new and overwrite your preexisting knowledge versus going with what you had before and making incremental changes? Part of the problem with going through massive… Part of the problem of things like if you’re trying to design an LLM or something like that, is, especially for English, there’s so many exceptions to the rules. If you want to rapidly learn the exceptions, you’re going to lose the rules, and if you want to keep the rules, you have a harder time learning the exception. David Marr is one of the early pioneers in computational neuroscience, and then Jay McClellan and my colleague, Randy O’Reilly, some other people like Neil Cohen, all these people started to come up with the idea that maybe that’s part of what we need. What the human brain is doing is we have this kind of actually a fairly dumb system, which just says, “This happened once at this point in time,” which we call episodic memory, so to speak. Then we have this knowledge that we’ve accumulated from our experiences of semantic memory. Now when we encounter a situation that’s surprising and violates all our previous expectations, what happens is that now we can form an episodic- … expectations. What happens is that now we can form an episodic memory here, and the next time we’re in a similar situation, boom. We can supplement our knowledge with this information from episodic memory and reason about what the right thing to do is. So it gives us this enormous amount of flexibility to stop on a dime and change, without having to erase everything we’ve already learned. And that solution is incredibly powerful, because it gives you the ability to learn from so much less information, really, and it gives you that flexibility. So one of the things I think that makes humans great is having both episodic and semantic memory. Now, can you build something like that? Computational neuroscience, people would say, “Well, yeah, you just record a moment and you just get it, and you’re done.” But when do you record that moment? How much do you record? What’s the information you prioritize and what’s the information you don’t? you just record a moment and you just get it, and you’re done.” But when do you record that moment? How much do you record? What’s the information you prioritize and what’s the information you don’t? These are the hard questions. When do you use episodic memory? When do you just throw it away? These are the hard questions we’re still trying to figure out in people. Then you start to think about all these mechanisms that we have in the brain for figuring out some of these things. And it’s not just one, but it’s many of them that are interacting with each other. And then you just take not only the episodic and the semantic, but then you start to take the motivational survival things, right? It’s just like the fight-or-flight responses that we associate with particular things, or the reward motivation that we associate with certain things, so forth. And those things are absent from AI. I frankly don’t know if we want it. I don’t necessarily want a self-motivated LLM, right? It’s like, and then there’s the problem of how do you even build the motivations that should guide a proper reinforcement learning kind of thing, for instance. So a friend of mine, Sam Gershman, I might be missing the quote exactly, but he basically said, “If I wanted to train a typical AI model to make me as much money as possible, first thing I might do is sell my house.” So it’s not even just about having one goal or one objective, but just having all these competing goals and objectives, and then things start to get really complicated.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.