CMOS mean field learning

Abstract

Analogue CMOS circuits which implement constrastive Hebbian learning have been designed, fabricated, and tested. A 19 neuron, 342 synapse mean field network, or deterministic Boltzmann machine (4.16 × 2.76 mm2 in 1.2 μm CMOS) has been constructed based on these synaptic circuits. Very small learning step size may be employed in this chip, which is advantageous for reliable mean field learning.