Energy efficiency in adaptive neural circuits

The human brain accomplishes its feat in cognitive general intelligence, unparalleled by any artificial intelligence to date even combining all of the world's current supercomputing power, with biological machinery that taken together are highly efficient and resilient, despite operating with fundamental neural and synaptic components that are intrinsically sluggish, noisy, and unreliable. Neuromorphic engineering [3] pursues naturally inspired artificial intelligence through exploiting physical foundations of computing with electronic devices and circuits that share similar thermodynamic properties and structural organization as ionic transport and neural circuits in the biological brain. While initial efforts in neuromorphic computing have shown modest success on the path towards general artificial intelligence, recent efforts in tandem with developments in machine learning using deep neural networks are targeting unprecedented performance levels in autonomy and efficiency of cognitive computing. Here we survey recent progress in such efforts towards 10-100x improvement in energy efficiency over the most advanced special-purpose cognitive computing platforms available today, through highly optimized neuromorphic integrated circuits that employ dense mixed-signal (analog and digital) neural and synaptic arrays with embedded highly flexible learning and inference function.