2 Comments
Jul 12Liked by xiq

Nice!! I've generally found myself a bit skeptical of the "free energy minimization" framework of cognition, as it has looked to me like its spread is powered as much by memetics as it is by true explanatory power. I also failed to understand the basic idea of why it would ever be a good strategy to tailor my actions so as it align reality with my predictions of it.

I found this article which you linked to be very helpful in starting to resolve some of this: https://latentobservations.substack.com/p/making-sense-of-active-inference

The high level idea almost looks to me like its saying that by doing planning in "observation space" instead of "model space," I can achieve some robustness to model misspecification. Which makes some intuitive sense!

What I haven't understood yet is how the framework addresses learning (which probably means I haven't understood all that much, haha). Maybe this is similar to your complaint about the prior being overly determinative... Definitely interested to spend some more time digging into a couple of these papers!

Expand full comment
author
Jul 13Author

Glad you got something out of it!

> how the framework addresses learning

My understanding is that learning and perception work exactly the same way, with learning affecting variables in the model that are causally upstream of the variables affected by perception..

Expand full comment