Analog Neuron as Generalised Similarity Metric

Human brain is well known to have unique ability to relate similarity between patterns. How it calculates the similarity with high accuracy is relatively an open ended problem. However, it is understood that trained neural networks create pathways in brain that can store encoded patterns as well as learn to distinguish between them. Inspired from this, we postulate that a single neuron itself could be used as a metric for emulating similarity calculation between patterns. The proposed neuron based similarity metric shows robust performance in pattern classification task even when different types of activation functions are used. The neural similarity circuit for pattern classification is implemented with a memristor crossbar - CMOS hybrid configuration. The system performance is tested for variability and robustness on wide range of practical conditions.