An Animated Study of Learning Dynamics in the 14-2-14 Encoder

This document accompanies a video of the same name originally presented at ICANN'93 (Am-sterdam) in September 1993. We use a modiied version of the Hyperplane Animator to visualize learning dynamics in the 14-2-14 encoder. The aim of the video is to demonstrate the eeects of common variations on standard Backpropagation (e.g., `soft' targets, on-line updates), alternative cost functions (Cross Entropy, Exception Error Function), alternative training algorithms (Conjugate Gradients, Quickprop), and a new output pattern encoding, Block Encoding, on the learning of the 14-2-14 encoder. The Hyperplane Animator provides a `window' onto the learning dynamics of the encoder task, allowing us to investigate the qualitative eeects of variations on the learning algorithm. In the case of the Exception Error Function and Block Encoding, the animations veriied our intuitions on how these modiications would innuence learning behaviour.