Event:

Abstract:
Take a moment to appreciate what you’ve done today: preparing for the day, commuting, writing emails, joking with colleagues, making dozens of small decisions - each involving prediction, error-correction, and adaptation. It’s easy to take for granted, but it’s a remarkable feat of inference.
Now consider what fueled it. A bowl of skyr, maybe a couple of eggs? That’s on the order of 12 watts of continuous metabolic power for your entire brain. Compare that to a modern AI assistant: to approximate a fraction of your ability, it had to read every piece of text ever written, 20 times over and use enough electricity to power a small country. To learn, modern AI uses back-propagation, an algorithm that requires storing and reusing activations across the entire network to compute a global error signal. Backpropagation, while powerful, is biologically implausible: it requires symmetric forward and backward weights, and a non-local error signal that has to traverse the entire network before any single unit can update.
The brain most certainly does not do backpropagation. A long tradition in cognitive science, captured by predictive coding and, more broadly, hierarchical Bayesian models of perception, proposes that the brain continuously generates top-down predictions about its sensory input, compares them against incoming signals, and updates its internal beliefs using only the resulting local prediction errors. Learning becomes a form of approximate Bayesian inference, distributed across a hierarchy, with no need for a global error signal or symmetric backward pathway.
In this talk, we present our work on making predictive coding networks competitive with back-propagation-trained networks by formalising them within this hierarchical Bayesian framework and, in doing so, moving closer to how brains learn: on the fly, continually, and on a fraction of the energy.

Join our Research Seminar

online:
Link: https://bbb.tum.de/rooms/sh4-fn5-pov-f3e/join


We are looking forward to a great discussion.

See full program: https://franziskaknolle.com/journal-club-tumnic/