Neuromorphics: Programmable analog computation through probabilistic digital communication
Abstract
Approach Lack of programmability is a major barrier to exploiting the efficiency of analog computation. Analog computation promises extreme energy-efficiency by operating close to the shot-noise limit. By exploiting physical laws (e.g., conservation of charge for summation), a handful of analog devices is sufficient to perform computation. In contrast, digital computation relies on abstractions that require many more devices to achieve the same function (e.g., hundreds of transistors to add two 8-bit numbers). Furthermore, these abstractions break when noise exceeds a critical level, requiring enormous noise margins to avoid catastrophic failures. In contrast, analog computation degrades gracefully, allowing for operation at low noise margins, thereby saving power. However, programmability requires flexibility, but this is difficult for analog computation because it exploits the underlying devices? fixed physical properties. PI has recently demonstrated that robust programmable computation could be realized with noisy heterogeneous components by combining analog computation with digital communication using a neural engineering framework inspired by the brain. The brain uses graded dendritic potentials (cf., analog computing), all-or-none axonal spikes (cf., digital communication) and probabilistically activated synapses (cf., connection weights). The analog computational units PI intends to use are spiking silicon neurons; and the digital communication fabric is a packet-switched network; and the connection weights are packet-delivery probabilities. While neuromorphic systems that combine some or all of these features of the brain have been built previously, they only performed specific computations. In contrast, this project realizes any desired mathematical computation by engineering the connections? weights to exploit the silicon neurons? heterogeneous tuning curves (the fabrication process introduces this variability). Objective Inspired by the brain’s energy efficiency, PI is exploring a hybrid analog-digital approach that uses subthreshold analog circuits to emulate graded dendritic activity and asynchronous digital circuits to emulate all-or-none axonal activity. Using this approach together with a formal method that maps arbitrary nonlinear dynamical systems onto spiking neural networks to develop a new breed of neuromoprhic chips that can be programmed to perform arbitrary computations. The goal of the project is to develop a programmable neuromorphic chip with a million neurons and a billion synaptic connections that consumes tens of milliwatts. When seamlessly interconnected by on-chip routers to build spiking neural networks with millions of silicon neurons and billions of synaptic connections, these chips offer a promising alternative for implementing intelligent controllers for miniature autonomous robots, an application that is extremely power-constrained.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512827
Entities
People
- Kwabena Adu Boahen
Organizations
- Office of Naval Research
- Stanford University
- United States Navy