Publications

We conduct a comprehensive empirical study of perturbation-based training algorithms, comparing their performance and properties across different network architectures and tasks.

Synaptic plasticity, the ability of synapses to change their strength, is a key mechanism underlying learning in biological neural networks. However, inferring plasticity rules from neural recordings is challenging due to the complexity of biological networks and the difficulty of measuring synaptic weights directly. Here, we introduce a model-based approach to infer plasticity rules from neural activity alone, without requiring direct measurement of synaptic weights.

We analyze the theoretical properties of node perturbation learning, providing new insights into its stability and scaling behavior in deep neural networks.

We develop a biologically plausible implementation of convolutional networks that maintains the computational advantages of weight sharing while respecting biological constraints on neural connectivity.