 |
Center for Nonlinear Studies |
Biologically inspired systems have revolutionized computing since the concept was pioneered by Carver Mead. Neuromorphic computing is about mimicking the neural structures as well as implementing competitive engineering practices, by understanding and exploiting the computational advantages of such structures. Moreover, analog signal processing techniques and bio-mimetic systems have been shown to be 1000 times more computationally energy efficient than their digital counterparts. In this talk, I will present my work on the design and implementation of such energy-efficient structures on hardware. I will discuss the design of fundamental blocks like the Hodgkin Huxley neurons and synapses, that remain true to the biophysics of neurons and perform compact, low-power, and real-time computation. To enable this computation and scale up from these modules, it is significantly important to have an infrastructure as well that adapts by taking care of the non-ideal effects like mismatches and variations, which commonly plague analog implementations. I will further discuss this tool framework and circuits to enable the built-in self-test algorithms for structures such as vector matrix multipliers. These cascade with programmable filters to implement core blocks used for analog signal processing systems. I will also discuss the algorithms built on hardware to solve linear systems of equations, taking an analog approach. By leveraging the strengths of silicon, these techniques provide several opportunities towards building energy-efficient systems.
Host: Anatoly Zlotnik