Distributional Active Inference

Hello. Welcome. It is March 5th, 2026. And we are live in active inference model stream number 21.1 with Bella Candemir, who will be presenting and discussing on distributional active inference. We'll have a presentation followed by a bit of discussion. So, if you're watching live, you can write questions in the live chat. Thank you for joining. To you for the presentation. >> Um thanks, Daniel, for the introduction. So, um the title of my talk is distributional active inference. I'm a machine learning researcher who is active mainly in doing like reinforcement learning research from a probabilistic perspective. So, a fun fact is that So, if I'm to give this presentation to a machine learning expert, so none of the three words you would see over here uh would be perceived in in their correct meaning. Uh at least I trust that so the active inference part is kind of what you will kind of understand correctly up front. Okay. So, let's make a thought experiment to set the stage. So, I'm sure quite many of the people in the audience know who Arthur Schopenhauer is. He is a famous philosopher. And he has a saying um which goes as follows. Like, one can indeed do what he wants, but he cannot want what he wants. So, that's the you know, direct translation of the original text. That's what he said around like 150 years ago or closer to 200, maybe. So, that what he points to is you know, you know, current lingo what the motor cortex is doing and what basal ganglia is doing. So, there are certain things we cannot keep control on or we don't feel like we're controlling. But, the thought experiment is as follows. So, what would happen if Schopenhauer lived today? And um you know, had access to all the recent advances in both AI and machine learning. As I both AI and computer computation neuroscience, sorry. So, he would probably append his saying in the following way. So, one can indeed do what she wants, probably what that's how he would frame it. And can also perceive what she wants. Um it's you know, pointing to the inside-out aspect of the brain. Uh but, she cannot want what she wants as he said before, neither can she perceive without predicting. So, prediction is essential to perceive stuff. And actually, we have a world model that kind of on which we have a control on kind of what to perceive. So, let's now imagine it that that was kind of a keynote speech he he he would give in in NeurIPS. So, um Why why why am I saying that? So, nowadays you know, machine learning especially reinforcement learning and computational neuroscience are in close touch. Reinforcement learning has already has always been kind of historically uh keeping a strong eye on computational neuroscience. And what I'm doing so, what I'm super interested in nowadays is to kind of tap all the kind of knowledge computational neuroscientists are developing on real let's say biological brains and use them in the machine learning framework as well as we can. So, how would I I like whole brain level cognition with my limited neuroscience background but a little stronger like machine learning background. So the brain has some functional areas at least broadly describable functional areas and they are in a kind of information flow. We have occipital cortex which where kind of sensory sensory stimuli are ending on. And so there is some information flow from here and there and there is kind of a central part of the brain. Where basal ganglia is located which is so to speak orchestrating the other parts. And the big questions are actually what kind of information is flowing from one area to the other? And what is actually the purpose of this flow from let say mechanistic point of view? So what kind of an energy function is being minimized? We can call it a Lyapunov function. And actually what kind of equations can we come up with to let say describe this flow mechanically? And so how on earth probably the most interesting one among the four to me. How on earth adaptive…