Reproduce “Normative Networks for Source Separation via Local Plasticity and Dendritic Computation” on OpenResearch

About this paper

Bariscan Bozkurt, Efe Ali Gorguner, Francesco Innocenti, Rafal Bogacz

The paper presents Predictive Entropy Maximization (PEM) for separating latent sources from unknown linear mixtures, including correlated sources. PEM approximates a log-determinant objective with output-variance rewards and normalized cross-covariance penalties, using prediction-error feedforward updates and covariance-based lateral interactions. In synthetic tests, it remains robust as correlation rises to 0.5 and tracks batch CorInfoMax under noise, without guaranteeing arbitrary mixtures.

Key insights

• A second-order Taylor approximation can effectively replace complex log-determinant objectives, enabling biologically plausible local dynamics for online determinant-based BSS.
• The resulting surrogate objective naturally promotes large component-wise variance while penalizing normalized cross-covariances, allowing for robust separation of correlated sources without strict independence assumptions.
• The proposed neural network architecture maps to local synaptic plasticity rules, including error-modulated Hebbian learning for feedforward weights and Hebbian-like updates for recurrent statistics, suitable for dendritic computation.

Summary by alphaXiv.