Neuromorphic Computing is a hardware-software co-design approach to realizing Artificial Intelligence (AI) by emulating brains.
Human brains can process real-time signals with remarkably low power consumption (about 20 Watts). Brains consist of a large number of neurons and synapses that operate in parallel (different regions) and communicate with each other through discrete spikes, leading to extraordinarily high computational efficiency. These spiking signals are referred to as Membrane Potential. The firing frequency of the membrane potentials in the nervous system is as low as ~ kilohertz level with millivolt-level magnitudes. Thus, a neuromorphic system can be ultra-high energy-efficient for large-scale artificial neural networks. Consequently, the neuromorphic system is capable of energy-restrained applications, such as edge computing, wearable devices, autonomous vehicles, unmanned aerial vehicles, spacecraft, and so on.
Additionally, biological neural systems have a self-learning capability referred to as associative memory and learning, allowing animals to deal with dynamic environments and plan optimal responses. Unlike prevailing deep learning relying on massive and labeled data, associative learning correlates concurrent events and memorizes the relationship between them. Consequently, the neuromorphic system with associative learning may eventually lead to self-learning intelligence.
Thus, our group aims for several specific topics: