Analog Hardware for Neural Computation

Our goal is to build a new type of analog computational system that uses less energy. The system is modeled after the brain and not simply a layer-after-layer architecture. The basic computational elements in this new architecture are multi-state and quasiperiodic oscillators. In a real brain the oscillatory elements are assembled into complex networks. This suggests two approaches to our research: (1) Construction and investigation of analog multi-state and quasiperiodic oscillator hardware with the ultimate goal of collecting state diagrams so a compiler can be written to exploit this hardware. These circuits can be built from analog components and run at lower power. (2) Assemble software simulations comprised of power-law distribution of nodes and edges into computational systems for demonstration of conventional and novel applications. We compare the ability of our oscillator circuits to store memory states with similar size Hopfield networks. That is, we will measure the number of stable oscillatory states as a function of the number of “neurons” and compare this with the number of fixed-point states in Hopfield networks for same numbers of “neurons.”

Team members:

Year 1 report | 12/2017 quarterly report References