Taro Toyoizumi, RIKEN

A Local Learning Rule for Principal Component Analysis and Independent Component Analysis

Humans can separately recognize individual sources when they sense a mixture of the sources in a noisy background. This computational process can be mathematically described by the combination of principal component analysis (PCA) for dimensionality reduction and independent component analysis (ICA) for separating independent sources. Here, we developed a biologically plausible unsupervised learning algorithm, error-gated Hebbian rule (EGHR), that performs the two using a single-layer neural network. The EGHR can extract the subspace that major principal components span similarly to Oja's subspace rule and can separate individual sources with negative kurtosis. Notably, it works when the conventional engineering approach that first applies PCA and second ICA fails due to high-variance noises. The EGHR can be easily implemented in a biological neural network or a neuromorphic chip because it updates synaptic strength based only on local information available for each synapse (i.e., a local learning rule). The results highlight the reliability and utility of the EGHR for large-scale parallel computation of PCA and ICA and its future implementation in a neuromorphic hardware.

Additional Information

Part of the Colloquium Series of the GRK 1589 "Sensory Computation in Neural Systems"

Organized by

Veronika Koren / Klaus Obermayer