This abstract presents the authors’ recent results on noise-enhanced computation, with specialized application to an error-correction decoding algorithm widely used in digital communication systems. Random defects are an increasingly troublesome problem for densely integrated electronic circuits. In this work we focus specifically on noise-induced transient upsets that can occur in nano-scale switching devices. There has recently been increasing interest in noise-tolerant design, since there is continual pressure to reduce the signal energy in electronic circuits, and eventually it will be necessary to perform computation at a very low signal-to-noise ratio. Some researchers propose sacrificing reliability in order to reduce energy consumption, which is acceptable for some applications that are inherently tolerant to momentary or intermittent faults (audio or image processing, for example). In contrast to noise-tolerant design, some researchers argue for noise enhanced solutions by using algorithms that benefit in some way from random upsets. Noise-enhanced solutions are partly motivated by biological examples, particularly neural signal processing, in which there are several instances where neural function is enabled or improved by some form of noise [3]. In this work, we present a case study showing how one algorithm is transformed into a noise-enhanced form. The transformation is not quite automatic, but nevertheless sheds light on a heuristic procedure that may be applied more broadly. The transformation is based on the stochastic gradient ascent heuristic, which is a well-known method for constrained optimization problems. The challenge is to transform the problem into a form suitable for applying the stochastic gradient ascent heuristic.
[1]
David J. C. MacKay,et al.
Information Theory, Inference, and Learning Algorithms
,
2004,
IEEE Transactions on Information Theory.
[2]
T. Wadayama,et al.
Gradient descent bit flipping algorithms for decoding LDPC codes
,
2008,
ISITA 2008.
[3]
Emmanuel Boutillon,et al.
Noisy Gradient Descent Bit-Flip Decoding for LDPC Codes
,
2014,
IEEE Transactions on Communications.
[4]
Shantanu Chakrabartty,et al.
Noise-Shaping Gradient Descent-Based Online Adaptation Algorithms for Digital Calibration of Analog Circuits
,
2013,
IEEE Transactions on Neural Networks and Learning Systems.
[5]
Mark D. McDonnell,et al.
The benefits of noise in neural systems: bridging theory and experiment
,
2011,
Nature Reviews Neuroscience.