On-Chip Learning of Hyper-Spectral Data for Real Time Target Recognition

摘要

It is well recognized that image based real time learning in neural computation using hyperspectral data set for target recognition is not only a complex problem to solve, but is also very time consuming. Solution of the same using neural processing with on-chip learning in hardware has never been attempted so far, even though such a high speed processing would be needed to perform real time data processing. It is important that a preprocessing step be included to cater to a high degree of input parallelism. Hence, a high-speed analog preprocessing method (developed and reported separately) that involved hardware-based data convolving scheme in a set of 3-D packaged chips with programmable templates, was used. Following that step, as the focus of our present paper, we have used the cascade error projection (CEP) learning algorithm (shown to be hardware-implementable) with on-chip learning (OCL) shceme to obtain three orders of magnitude speed-up in target recognition compared to software-based learning schemes. Thus, it is shown, real time learning as well as data processing for target recognition can be achieved. This paper first describes the processing details of a hyper-spectral data set in software containing a large 356,000 data points image. The paper also describes the fabrication of a test chip that was caharacterized and evaluated for comparison of its performance with the emulation results reaching up to 96.9percent correct target recognition.