Neuromorphic electronic systems

Abstract

Biological in forma tion TWO TECHNOLOGIESHistorically, the cost of computation has been directly related to the energy used in that computation. Today's electronic wristwatch does far more computation than the Eniac did when it was built. It is not the computation itself that costs-it is the energy consumed, and the system overhead required to supply that energy and to get rid of the heat: the boxes, the connectors, the circuit boards, the power supply, the fans, all of the superstructure that makes the system work. As the technology has evolved, it has always moved in the direction of lower energy per unitcomputation. That trend took us from vacuum tubes to transisitors, and from transistors to integrated circuits. It was the force behind the transition from n-MOS to CMOS technology that happened less than ten years ago. Today, it still is pushing us down to submicron sizes in semiconductor technology.So it pays to look at just how much capability the nervous system has in computation.There is a myth that the nervous system is slow, is built out of slimy stuff, uses ions instead of electrons, and is therefore ineffective. When the Whirlwind computer was first built back at M.I.T., they made a movie about it, which was called "Faster than Thought." The Whirwind did less computation than your wristwatch does. We have evolved by a factor of about 10 million in the cost of computation since the Whirlwind. Yet we still cannot begin to do the simplest computations that can be done by the brains of insects, let alone handle the tasks routinely performed by the brains of humans. So we have finally come to the point where we can see what is difficult and what is easy. Multiplying numbers to balance a bank account is not that difficult. What is difficult is processing the poorly conditioned sensory information that comes in through the lens of an eye or through the eardrum.A typical microprocessor does about 10 million operations/s, and uses about 1 W. In round numbers, it cost us about l O -' J to do one operation, the way we do it today, on a single chip. If we go off the chip to the box level, a whole computer uses about 10-5J/operation. Awhole computer is thus about two orders of magnitude less efficient than is a single chip.Back in the late 1960's we analyzed what would limit the electronic device technology as we know it; those calculations have held up quite well to the present [I]. The standard i ntegrated-ci rcu it fabricat ion processes available today allow usto build transistorsthat have minimum dimensions of about 1 p m). By ten years from now, we will have reduced these dimensions by another factor of 10, and we will be getting close to the fundamental physical limits: if we make the devices any smaller, they will stop working. It is conceiveable that a whole new class of devices will be invented-devices that are not subject to the same limitations. But certainly the ones we have thought of up to now-including the superconducting ones-will not make our circuits more than abouttwoordersof magnitude more den...

Probabilistic synaptic weighting in a reconfigurable network of VLSI integrate-and-fire neurons

“…Modern complementary metal-oxide-semiconductor (CMOS) very large-scale integration (VLSI) technology provides us with a medium that is well-suited to the implementation of neural systems. MOS devices can operate in regimes where their physics are identical to that of key neural structures, and VLSI technology enables extremely dense integration of these devices (Mead, 1990). We believe that the coupling of this technology with neurally inspired architectures will allow us to exploit the potential of spike coding schemes, and ultimately enable us to close the performance gap.…”

Computation in Neuromorphic Analog VLSI Systems

“…The term "neuromorphic" was coined by Carver Mead to describe very large scale integration (VLSI) systems containing electronic analog circuits that mimic neurobiological architectures present in the nervous system [18]. Neuromorphic computation is related to modeling and simulation of networks of neurons and systems using the same organizing principles found in real nervous system.…”

Reconstruction, Identification and Implementation Methods for Spiking Neural Circuits

“…This thesis addresses a number of information processing and modelling problems relevant to the field of computational neuroscience as well as neuromorphic systems engineering (Mead, 1990). …”