Wisard

Wisard, often written WiSARD in some sources, is a family of weightless neural networks designed for pattern recognition and classification. Unlike traditional neural networks, Wisard uses RAM-based neurons and stores training examples directly in memory without adjustable synaptic weights. The architecture consists of multiple discriminators, one per target class. Each discriminator is built from several RAM blocks; each RAM receives a binary input pattern (or a binarized version of the input) and acts as a simple memory lookup. Training consists of presenting binary representations of examples to the corresponding class discriminators, which populate the RAM addresses. Recognition proceeds by feeding a binary input to all discriminators; each RAM block outputs a signal if the input matches stored addresses, and the discriminator aggregates responses to vote for its class. The final decision is typically taken by selecting the class with the strongest consensus among discriminators, often via a threshold or majority rule.

Wisard systems are noted for fast training and straightforward hardware implementation, making them suitable for real-time and embedded applications, particularly when data can be binarized. They are robust to some noise due to their distributed memory structure but can be sensitive to how the input is discretized and may require careful design of RAM block sizes and address schemes. Applications historically include optical character recognition, symbol recognition, and other small-scale pattern recognition tasks. See also weightless neural networks, RAM-based neural networks.