Activation
Error Criterion
Gradient Search
Input
Probe

Locally-Connected Synapse DLL

A common problem with using a fully-connected neural network for image processing problems is that even a modest sized image requires an enormous number of weights. One way to solve this problem is to replace the fully-connected matrix of weights at the first layer with one that is only locally-connected. For instance, if the input image is 400x400 pixels and the first hidden layer is a 40x40 PE Axon, then each PE of the hidden layer would be fed by a 10x10 matrix of weighted connections from the input. In this way, much of the spatial information of the image is preserved while the number of weights is drastically reduced (from 256,000,000 for the fully-connected case down to 160,000 for the locally-connected case).

Download zip file (0.1M)