Neuromorphic Computing

By Rebooting Computing Portal Staff

All conventional computers are based on the classic von Neumann architecture ( going back to the first vacuum-tube computers from around 1950. These have a separate arithmetic logic unit (ALU) and a memory unit, with data being shuttled between them. This is quite different from the device architecture in biological computers, i.e., brains, where logic and memory are closely integrated in the same basic device, the neuron, and in connections between neurons, known as synapses. A neuron combines the outputs of many other neurons (up to thousands) via addition or subtraction in generating its own output. The strength and polaraity of these synapse connections constitute memories, which may change over time owing to patterns of use, enabling adaptation of both memory and logic. This enables neuron circuits to be particularly good at parallel processing for pattern matching and memory retrieval. Neuromorphic circuitss may be based on conventional transistors, or alternatively novel devices such as memristors ( or Josephson junctions.

A recent article in MIT Technology Review describes recent research efforts in neuromorphic computing ("Thinking in Silicon"), and features an ongoing research program sponsored by the US Defense Advanced Research Projects Agency (DARPA,, with major projects at IBM and HRL. The IBM project is developing systems consisting of one million silicon neurons (each based on circuits of several thousand transistors), and has already demonstrated machine learning for image processing and game playing on a smaller brain model. For further information, see A. Cassidy, et al. (IBM), "Cognitive Computing Building Block: A Versatile and Efficient Digital Neuron Model for Neurosynaptic Cores", . The HRL group is developing a neuromorphic chip that combines memristors with conventional CMOS circuits (see for more information), and is targeting a lightweight mobile system such as a miniature aircraft. These efforts are aiming not to replace conventional number-crunching computers, but to complement them for applications in artificial intelligence where the neuromorphic approach may be more compact and efficient.

Another recent extensive review of the field is "Finding a Roadmap to Achieve Large Neuromorphic Hardware Systems", by Jennifer Hasler and Bo Marr, available online at