Brain-like variational inference

Hello, welcome everyone. It is June 25th, 2025. We are in Active Inference Guest Stream 112.1, Brain-like Variational Inference with Hadee Fafai. And I looked and found that actually model stream 11.1 when we discussed Poisson Variational Autoencoder was on June 26th, 2024. So just kind of funny that it's like exactly one year or all but a day away from a year. So it's awesome to check back in with you and hear these research updates. So thanks again for joining and looking forward to the presentation. All right, awesome. Thanks, Daniel, for the invitation. I had a great time last time when I presented and I hope we can have fun this time again. All right, so the work that Daniel mentioned, the Poisson, it's I think this one. So you can check it out. It's from last year and it is published in NeurIPS last year. And this is the new preprint, brain-like variational inference, which we're going to talk about today. And that's a work in collaboration with Dekel, who's an ex PhD student in Berkeley, and Jake, who's my postdoc mentor. All right, but you know, in the title, there's like several words, brain-like variational inference. And I hope that I can communicate what each of these words mean to you by the end of this talk. And let's first start by inference. What is inference? It's understanding the world. So our brains receive sensory information from the world out there. And we have to make sense of the state of the world. We have to understand how the world is right now so that we can act in it and survive. And in fact, this is the topic of a new blog post that I have written. I'm going to plug it here. So this blog post is coming soon. It's probably in a week on my website, mysterioustoon.com. And inference is, you know, I claim that it's your brain trying to guess the hidden cause of its observations. And, you know, to just drive this point home, I have, I give you three very concrete examples, each more difficult than the other. For example, if you go outside and see there is extensive wetness of the ground, you might guess that, you know, there, there was rain last night. So the cause of what you observe, which is wet ground was rain. So that's inference, you're actually doing inference when you reach that conclusion. It's a very simple kind of inference. Also, I discuss this problem of degeneracy in visual perception. When you see a circular shape, it could be a 2d circle, you know, held perfectly perpendicular to your direction of gaze, or it could be actually a 3d sphere. And, you know, intuitively, you might think that it's most likely a sphere, but where does that come from? That's your prior knowledge. So having lived in this world, we have seen more spheres than circles. Therefore, you might actually intuitively think that what you're seeing is actually a sphere. And last one is the most difficult, nearly impossible type of inference is the social inference. So when somebody says something to you, you want to understand what they actually meant. And that is also can be understood as an instance of inference. In linguistics, this is technically called pragmatics. But I claim and develop an argument that all of these examples are, you're basically performing inference. And so I'm going to post this blog post in like a week or so. And you can either subscribe to the website or follow me on social media like Twitter and LinkedIn, I'm going to post it there too. Here's another example from the blog post. Most of us here probably are familiar with this allegory of the cave. So Plato argued that being alive is just like, you know, being prisoners, but in the back of a cave, you know, you're separated from the real world out there. This is the real world. You don't really have access to it. But you see shadows from the real world projected on the wall that you have access to. And you know, when you see those shadows or observations, you perform inference to understand the likely causes that caused those observations.…