14. Artificial Synapses for Learning
Category: Brain-like Computing
Paper 17.1 - NVM Neuromorphic Core with 64k-cell (256-by-256) Phase Change Memory Synaptic Array with On-Chip Neuron Circuits for Continuous In-Situ Learning; S. Kim et al, IBM
Advances in machine learning and neuroscience have sparked growing interest in neuromorphic (brain-inspired) computing, and a number of neuromorphic circuits have been demonstrated that are capable of functions such as pattern-recognition. At the IEDM, IBM researchers will describe a chip that may bring neuromorphic computing closer to true artificial intelligence: the largest neuromorphic “core” ever built, a 256 x 256 array of artificial synapses with on-chip programming circuitry. It may be capable of “deep learning,” which is when machines follow sophisticated algorithms in an attempt to mimic brain functions like seeing, listening and thinking. The synapses are 64k-cell phase-change memory (PCM) devices, and the researchers say that each PCM synapse is capable of running in one of three modes independently, each of which is an analog of the behavior of real neurons: 1) so-called leaky-integration-and-fire (the synapse fires when input voltage reaches a certain threshold); 2) spike-timing dependent plasticity (an algorithm that mimics a fundamental brain mechanism for learning and memory; and 3) in idle mode. The researchers say that once wiring issues are solved the array size potentially could be increased to the biological scale.