Algorithms and Applications for Cognitive Computing Systems
Abstract
Toward a new era of cognitive computing, IBM unveiled a trilogy of innovations on TrueNorth system that was inspired by the unparalleled functional variety and efficiency of human brains. The non-von Neumann nature of the TrueNorth architecture proposes a novel approach to realize many neuromorphic algorithms and applications. However, conventional neuromorphic computing models are not necessarily efficient when running on the TrueNorth systems for its spiking-based computation process. For specific applications, creating a complete specification of neuro-synaptic cores as well as a generic format of the inputs and outputs that are compatible to the TrueNorth architecture is challenging with the increase of the network size and the complexity of the problems. A new way of thinking and developing the algorithms and applications that are intrinsically efficient and scalable on TrueNorth architecture becomes very crucial for the successful adoption of TrueNorth systems. The objectives of this project include 1) to systematically investigate the methodologies of realizing general inference models such as recurrent neural network (RNN) and convolutional neural network (CNN) on TrueNorth architecture, especially considering the spiking based interface and computation process; and 2) to holistically explore the development and optimization of image classification models like CAFFE and other applications, such as video anomaly detector, etc. on top of the TrueNorth architecture. Detailed performance and power evaluations of the developed models and applications on TrueNorth architecture will be also performed and compared with the counterparts on CPU, GPCPU, and FPGA.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 11, 2016
- Source ID
- FA87501510176
Entities
People
- Yiran Chen
Organizations
- Rome Laboratory
- United States Air Force
- University of Pittsburgh