Ultra low power analog design and technology for artificial neurons

Abstract

In a context of the end of Moore's law, energy dissipation constitutes a real challenge. Among the new energy efficient paradigms for data processing, bio-inspired computing is very promising, moreover introducing cognitive characteristics. As applications at very high scale are addressed, the size and energy dissipation of both the neuron and synapse cells needs to be minimized. In this context, this paper presents the design of an original artificial neuron, using standard 65nm CMOS technology with optimized energy efficiency and its application in basic neural networks. By recalling brain and neuron features, it is shown why neuron energy efficiency is roughly limited to 1 pJ/spike in biological neuron. Biological and artificial neurons features are carefully compared, highlighting the importance of downscaling. The artificial neuron circuit presented was designed to exhibit wide band spiking frequencies, targeting large scale bio-inspired information processing applications. The most important feature of the fabricated circuits is the neuron energy efficiency in the few fJ/spike range, which improves prior state-of-the-art by two to three orders of magnitude. This performance is achieved by minimizing two key parameters: the supply voltage and the related membrane capacitance. Meanwhile, the obtained standby power at a resting output does not exceed tens of picowatts. The circuit is sized to 35μm2, reaching a spiking output frequency of 26kHz. It is then shown how this artificial technology has already been used for two applications: (i) emulation of bursting mode, important in brain stimulation context or for robotics (locomotion rhythm), (ii) stochastic resonance application, useful to detect an electrical signal buried in noise, phenomena exploited in nature by species. These results already allow envisioning the development of highly integrated neuro-processors (vision application). A variant circuit (biomimetic) could be used for robotics, neuroscience or medical applications.