Lead Inventor:
Aurel A. Lazar, Ph.D.
A recurrent neural network implemented using highly-parallelized analog circuits for recovering time-encoded visual signals, mimicking the properties of biological neurons
Time encoding can faithfully and efficiently capture visual information in the time domain using a series of spikes, in a manner mirroring biological sensory systems. However, as visual signals are necessarily high bandwidth and must be represented using a significant number of measurements, recovery of time-encoded visual signals using conventional computationally-intensive numerical methods require an inordinate amount of time. By structuring the recovery of the time-encoded signal as an optimization problem, the signal can be efficiently decoded using a recurrent neural network. This technology implements a video time decoding machine using a highly-parallelized VLSI analog circuit structured as a recurrent neural network, allowing for real-time recovery of the video without the need for a computer.
Time-encoded video signals can be recovered using parallel analog circuits without the need for complex numerical algorithms and a separate computer.
Previously, in order to recover time-encoded signals, complex numerical methods requiring significant time-consuming computations were used. Realizing a recurrent neural network-based decoder in analog hardware not only enables fast real-time recovery of time-encoded signals without the need for a computer, but also practical implementation of this technology based on its use of standard silicon processes.
The analog circuit implementation of the decoder has been evaluated via computer simulation verifying the quality of the reconstructed video signal in addition to its speed.
Applications:
-- Enables practical video time-encoding and decoding in low-voltage analog systems
-- High-resolution, low-voltage analog-to-digital converters (ADCs)
-- Mirrors biological representation of visual stimuli -- information is represented as a sequence of temporally-varying action potentials -- lending to applications in computer vision
Advantages:
-- Time domain signal representations can exploit increasing time resolution of highly scaled integrated analog circuits
-- Does not require a computer to perform slow and complex numerical algorithm
-- Can be realized as a highly-parallelized analog VLSI circuit
-- Mimics the properties of biological visual systems
Patent Information:
Patent Issued
Licensing Status: Available for licensing and sponsored research support
Related Publications:
Lazar, A.A.; Yiyin Zhou; , "Realizing Video Time Decoding Machines with recurrent neural networks," Neural Networks (IJCNN), The 2011 International Joint Conference on , vol., no., pp.1027-1034, July 31 2011-Aug. 5 2011
Aurel A. Lazar, Yiyin Zhou, Massively parallel neural encoding and decoding of visual stimuli, Neural Networks, Available online 16 February 2012, ISSN 0893-6080, 10.1016/j.neunet.2012.02.007.
