Exponential Capacity in an Autoencoder Neural Network with a Hidden Layer

Abstract: A fundamental aspect of limitations in learning any computation in neural architectures is characterizing their optimal capacities. An important, widely-used neural architecture is known as autoencoders where the network reconstructs the input at the output layer via a representation at a hidden layer. Even though capacities of several neural architectures have been addressed using statistical physics methods, the capacity of autoencoder neural networks is not well-explored. Here, we analytically show that an autoencoder network of binary neurons with a hidden layer can achieve a capacity that grows exponentially with network size. The network has fixed random weights encoding a set of dense input patterns into a dense, expanded (or \emph{overcomplete}) hidden layer representation. A set of learnable weights decodes the input patters at the output layer. We perform a mean-field approximation of the model to reduce the model to a perceptron problem with an input-output dependency. Carrying out Gardner's \emph{replica} calculation, we show that as the expansion ratio, defined as the number of hidden units over the number of input units, increases, the autoencoding capacity grows exponentially even when the sparseness or the coding level of the hidden layer representation is changed. The replica-symmetric solution is locally stable and is in good agreement with simulation results obtained using a local learning rule. In addition, the degree of symmetry between the encoding and decoding weights monotonically increases with the expansion ratio.