TOTEM: A HIGHLY PARALLEL CHIP FOR TRIGGERING APPLICATIONS WITH INDUCTIVE LEARNING BASED ON THE REACTIVE TABU SEARCH

Abstract

The training of a Multi-Layer Perceptron (MLP) classifier is considered as a Combinatorial Optimization task and solved using the Reactive Tabu Search (RTS) method. RTS needs only forward passes (no derivatives) and does not require high precision network parameters. TOTEM, a special-purpose VLSI chip, was developed to take advantage of the limited memory and processing requirements of RTS: the final system effects a very close match between hardware and training algorithm. The RTS algorithm and the design of TOTEM are discussed, together with the operational characteristics of the VLSI chip and some preliminary training and generalization tests on triggering tasks.