In the brain, synaptic learning is constrained to use spatio-temporal local information. However, the majority of the learning algorithms that we use to train neural networks violate spatial or temporal locality (such as RTRL and BPTT).
How can the brain efficiently assign credit using only local spatio-temporal information?
In this discussion group, we aim to explore biologically plausible solutions to the spatio-temporal credit assignment problem. Additionally, I will introduce our computational framework (GLE) for fully local spatio-temporal credit assignment in physical, dynamical networks of neurons. This framework approximates BPTT in deep cortical networks with continuous-time neuronal dynamics and continuously active, local synaptic plasticity.
Reference
Ellenberger, B., Haider, P., Jordan, J., Max, K., Jaras, I., Kriener, L., Benitez, F., & Petrovici, M.A. (2024). Backpropagation through space, time, and the brain. ArXiv, abs/2403.16933.
Go to group wiki Go to wiki users Info
| Day | Time | Location |
|---|---|---|
| Fri, 03.05.2024 | 16:00 - 18:00 | Panorama |