Initializing receptive field-like weights in BP learning network (II) - Internal representation

摘要

With the T-C recognition task in BP neural network finished, the final distributions of weights between input and hidden layer have been investigated under the conditions of receptive field(RF)-like and random initializing in global or local connective networks. It is discovered that the final distributions of weights W show regular global symmetric distribution under RF-like initializing regardless of global and local connective networks; the distribution of central hidden units is similar to concentric RF-like distribution. But in the experiment on random initializing, this regularity did not appear, and the final distributions of weights are full of variety: some distributions of hidden units resemble concentric RF-like distribution in retina andLGN or asymmetric RF in visual cortex. These phenomena exist in multi-channel BP networks.