 |
Center for Nonlinear Studies |
The end of Moore’s Law is rapidly transforming the computing. Quantum annealers and neuromorphic systems are next generation processors which are offer advances in computational ability and energy efficiency. There are several challenges in pre-processing, and algorithm design which affect how these processors can be incorporated into heterogeneous workflows. In this seminar, I will present several graphical elements that are incorporated into workflow or algorithm design for quantum annealers and neuromorphic hardware. Quantum annealers can solve optimization problems posed as Ising spin glasses by finding the global minimum of the energy landscape. This search can be aided through quantum tunneling effects, which allow the system to escape local minima. Defining a spin system, and finding its embedding onto the hardware, are difficult problems to solve. I will present results for D-Wave’s 2X annealer intended to expedite the minor embedding step by utilizing a “look up” table approach indexed by a finite set of graph minors. Neuromorphic processors are event-driven computing platforms which execute programs through the transmission of discrete time signals. They have had widespread success in efficient, low-power deployment of deep neural networks such as convolutional neural networks, deep belief networks and recurrent neural networks. However, training and constructing these networks require substantial computing resources and create a significant bottleneck in the development of heterogeneous computing workflows. Utilizing graph theoretical and spin-glass physics concepts, I will describe an approach to algorithmic design explores the feasibility of using neuromorphic architecture to implement graph algorithms, while avoiding the need for extensive training resources. In particular I will demonstrate how spiking neural systems can be constructed and used to implement a spike-based label propagation method for community detection in undirected, unweighted graphs. This approach will be derived from the nonlinear dynamics of correlated spiking neurons and then adapted for deployment on IBM’s TrueNorth Neurosynaptic System.
Host: Hristo Djidjev