Independent Component Analysis in Neuron Model

Independent Component Analysis (ICA) and related algorithms provide a model for explaining how sensory information is encoded in our brains. However, it remains unclear how neural network uses its biological plasticity to achieve this ICA-like processing: maximization of information transmission or minimization of redundancy. Here, we consider a neuron model proposed by Savin, Joshi, and Triesch (2010), which includes three forms of plasticity in real neural network: spike-timing dependent plasticity (STDP), intrinsic plasticity (IP), and synaptic scaling. We investigate both theoretical and experimental aspects of the model and found that the three types of plasticity play important but different roles in efficiency and quality of learning. Although this neuron model cannot compete with classic ICA algorithms in solving blind separation problem, it provides a biological perspective that can potentially explain how our brains learn and why our brains have such high capacity and complexity.