3. LeNet CNN Architecture

This Wiki page provides an overview of the LeNet architecture.

LeNet CNN Architecture

The models package of WatChMaL/CNN contains the kazunet module, which is an implementation of LeNet with small adjustments in filter sizes and number of convolution layers. LeNet was one of the first convolutional neural network architectures to be designed as a replacement for hand-programmed feature extractors. It consists of a series of convolution layers connected via sub-sampling connections connected to three fully-connected layers at the output (see below).

Unlike the original LeNet-5, kazunet uses max-pooling as a down-sampling scheme between convolutional connections and employs the ReLU in place of the sigmoidal activation function used in LeNet.