Learning Interpretable Hidden State Structures for Handwritten Numeral Recognition

A novel paradigm of learning interpretable hidden state structures from the Multi-Layer Perceptron (MLP) neural network is proposed, in this work, that considers hidden state activations as the feature vector. The classification task at hand is handwritten numeral recognition. The k-means Elbow method is used to analyze the interpretability of the clusters formed from each hidden layer. Only the hidden layer that forms well-defined and linearly separable clusters, corresponding to the different numeral classes, is considered for the feature extraction. An interpretable structure thus emerges from the hidden representations that are learnt using a suitable classifier. Direct learning of hidden states using the linear k-Nearest Neighbor (kNN) classifier and the Support Vector Machine (SVM) with linear kernel yields high accuracies. Two different MLP configurations are tested for the purpose. Patch-based hidden structure learning with feature fusion is found to further improve the results. Experiments on the benchmark MNIST handwritten numeral image dataset prove the efficacy of our interpretability-based approach as compared to existing works.