Artificial Neural Network: Deep or Broad? An Empirical Study

摘要

Advent of Deep Learning and the emergence of Big Data has led to renewed interests in the study of Artificial Neural Networks (ANN). An ANN is a highly effective classifier that is capable of learning both linear and non-linear boundaries. The number of hidden layers and the number of nodes in each hidden layer (along with many other parameters) in an ANN, is considered to be a model selection problem. With success of deep learning especially on big datasets, there is a prevalent belief in machine learning community that a deep model (that is a model with many number of hidden layers) is preferable. However, this belies earlier theorems proved for ANN that only a single hidden layer (with multiple nodes) is capable of learning any arbitrary function, i.e., a shallow broad ANN. This raises the question of whether one should build a deep network or go for a broad network. In this paper, we do a systematic study of depth and breadth of an ANN in terms of its accuracy (0-1 Loss), bias, variance and convergence performance on 72 standard UCI datasets and we argue that broad ANN has better overall performance than deep ANN.