Letters of the Alphabet: Discovering Natural Feature Sets (2202.10934v2)

Abstract: Deep learning networks find intricate features in large datasets using the backpropagation algorithm. This algorithm repeatedly adjusts the network connections.' weights and examining the "hidden" nodes behavior between the input and output layer provides better insight into how neural networks create feature representations. Experiments built on each other show that activity differences computed within a layer can guide learning. A simple neural network is used, which includes a data set comprised of the alphabet letters, where each letter forms 81 input nodes comprised of 0 and 1s and a single hidden layer and an output layer. The first experiment explains how the hidden layers in this simple neural network represent the input data's features. The second experiment attempts to reverse-engineer the neural network to find the alphabet's natural feature sets. As the network interprets features, we can understand how it derives the natural feature sets for a given data. This understanding is essential to delve deeper into deep generative models, such as Boltzmann machines. Deep generative models are a class of unsupervised deep learning algorithms. The primary function of deep generative models is to find the natural feature sets for a given data set.