Principal Investigator: Professor Yang Hyunsoo, ECE

Recently, significant research efforts have been made to find alternative computing architectures than traditional von Neumann one, where the transfer of data between physically separate logic and memory blocks results in processing bottlenecks and significant power consumption. We aim to demonstrate a real-time neural-like computing system implemented using ultra-low power stochastic magnetic tunnel junctions (MTJ). It is a novel computing system which uses neural-like devices for computing and non-volatile magnetic devices for storing. This architecture provides substantially low power consumption and fast calculation by physically emulating neurons at the very low device and circuit level.

As a final demonstration, we will apply such a computing system in mimicking a brain’s function, in particular, for an audio signal separation task. Audio signal separation has been one of the most difficult tasks in the audio signal processing area with a great potential in speech recognition, simultaneous interpretation, chatting robots, security monitoring, self-driving car and so on. Although advances in audio signal separation have led to a performance improvement in challenging scenarios such as noisy and far-field conditions using various artificial neural networks (ANNs), it still performs poorly when the signal source of interest is recorded in crowded environments and when the number of signal sources increases. We aim to do a real-time separation of the audio signals of a target source from multi-source signals with a low latency, power consumption, and error rate.