Explainable Feature Extraction Using a Neural Network with non-Synaptic Memory for Hand-Written Digit Classification

The human brain recognizes hand-written digits by extracting the features from a few training samples that compose the digit image including horizontal, vertical, and orthogonal lines as well as full or semi-circles. In this study, we present a novel brain-inspired method to extract such features from handwritten digits images in the MNIST database (Modified National Institute of Standards and Technology database). In this study, we developed an explainable feature extraction method for hand written digit classification in which the extracted information are stored inside the neurons as non-synaptic memory manner. For this purpose, a neural network with 10 single neurons was trained to extract features of training images (each neuron represents one digit class). Following that, the trained single neurons are used for the retrieval of information from test images in order to assign them to digit categories. The accuracy of the classification method of test set images is calculated for different number of training samples per digit. The method demonstrates 75 % accuracy using 0.016 % of the training data and maximally shows accuracy 86 % using one epoch of whole training data. The method as an understandable feature extraction method allows users to see how it works and why it does not perform well on some digit classes. To our knowledge, this is the first model that stores information inside single neurons (i.e., non-synaptic memory) instead of storing the information in synapses of connected layers. Due to enabling single neurons to compute individually, it is expected that such class of neural networks show higher performance compared to traditional neural networks used in complicated classification problems.