Biologically Plausible Learning with Non-Negative Activity and Fixed-Sign Synapses
A recent preprint on arXiv (2608.06963) presents a neural architecture grounded in biological principles that adheres to Dale's law, which asserts that neurons are exclusively excitatory or inhibitory, and synapses maintain a consistent sign. The innovative model employs two non-negative channels to capture both positive and negative influences, drawing inspiration from the brain's on-off signaling. This approach facilitates learning akin to backpropagation while avoiding mixed-sign values. It tackles a significant hurdle in developing biologically plausible learning frameworks, which frequently breach this fundamental characteristic of cortical networks. This research holds significance at the crossroads of neuroscience and artificial intelligence, potentially leading to more brain-inspired learning methods.
Key facts
- The paper is available on arXiv with ID 2608.06963.
- It addresses Dale's constraint, which prohibits neurons from being both excitatory and inhibitory.
- The proposed architecture uses non-negative activity for neural activations and learning signals.
- Synapses have fixed sign in the model.
- The approach uses two complementary non-negative channels for positive and negative contributions.
- It is inspired by evidence of on-off representations in the brain.
- The model supports backpropagation-like learning.
- The work aims to bridge biologically plausible learning and deep learning.
Entities
Institutions
- arXiv