Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hello, greetings. It's November 11th, 2022. We're in Cohort 1 of the textbook group, and we're coming back to Chapter 10, looking at the sections of 10, asking any questions about 10, thinking about any overview points. Before next week, when it will be our final meeting, and we'll talk about feedback, fill out the form in future textbook groups, talk about project ideas, next steps, talk about our plans for 2023, and so on. But where do we begin with 10? We covered many of the sections once last week. There's a lot of sections. The chapter kind of is like a symphonic presentation. It has multiple movements and subsections that have their own rhythm and structure. Let's go to the end. Let's go to the end. I think this is a really important note about learning and active engagements, which is if you're not, even before skin in the game and psychological ownership and so on, if one is not making predictions, not just about material, like what's going to happen next in the TV show that I'm watching, oh, it surprised me, but about the consequences of one's actions in the world, they will not be able to update those models of action in the world. So you could be surprised and end up being a good inferor in a passive inference setting. However, doing is what actually allows you to learn in a totally different way, because whether you realize it at the beginning of the journey or not, or whether you even agree with this or not, those actions that you take are chosen and exist on a trade-off frontier of epistemic and pragmatic value. And so simply choosing to act is a strategy that opens up a space for pragmatic as well as epistemic action. It's kind of like the hyper-prior-on-active inference is how active. And one other note, just that anyone can add anything. This was a discussion with JF and Candon some days ago in our research meeting about the scaling debate in machine learning. All you need is scaling. Bigger natural language models and so on. And like, how can active inference and how can we contextualize that and improve the state of things? And we talked about how, importantly, GPT does not choose its own inputs. And so, yes, it can engage within a perception action loop related to receiving inputs and sending outputs. However, in terms of its own training and development, it's importantly constrained. And that's not necessarily a bad thing. It's just that to actively select what stimuli to sample is very central to active inference. From the ocular motor to the tactile. And so, all of the big data training algorithms do not select their own inputs. And then if they were to, it could be done in two ways. It could be done in an ad hoc way or principled way. The ad hoc way would be you do some kind of different architecture of a model that, you know, predicts what inputs would be valuable to train. or you could think of a unified imperative that would decide on both action and learning rules. Eric and Brock. Yeah, I suppose you could say that reinforcement learning, the machine is selecting its own samples to get feedback on. Now, reinforcement learning is not GPT-3. It's not language learning, language modeling. but I think an earlier paradigm, maybe I don't know. I think they still use reinforcement learning and deep learning models. Although, you know, it was also, I think a lot of people are discouraged by that because it's, it, you know, as they say, it's a, it's a very narrow, a very low bandwidth signal you're getting back. But it is an example of learning models using, generating their own samples. Thanks, Brock. Um, I was gonna, I guess, uh, take it to this biological place we left off in last session of the circadian of like, being a kind of data cleaning process almost, or, um, stabilizing the, you know, the like data set coming in. Um, and, um, that we do have some very, very crude, like, uh, you know, built-in labels for edges kind of in our visual cortex, right? Uh, that are presumably evolutionary. Um, and some…