Hammerstrom, D. 2015, Darpa neurocomputing. Paper presented at 2015 IEEE International Electron Devices Meeting (IEDM).
|
|
|
Abstract
|
DARPA is investigating machine learning algorithms and computer architectures that mimic selected characteristics of human intelligence, such as learning and pattern recognition, to address the challenges of data recognition, control and complexity in the continuously evolving environments where DoD systems operate. Learning approaches to date have shown great promise in solving a wider range of problems in less constrained environments, but require high-precision and long compute times, limiting their ability to learn large data sets rapidly or adapt in real time in the field. Neural inspired algorithms allow the use of low precision, hierarchical, temporal memory structures that can rapidly evolve with changing data, minimizing the need for long training times and maximizing rapid, on-line (in the application) real time adaptation. These capabilities coupled with optimized silicon will result in high performance and low power for real-time, embedded system operation for a wide range of applications.
|
