CMOS Circuits for Shape-Based Analog Machine Learning
share
While analog computing is attractive for implementing machine learning (ML) processors, the paradigm requires chip-in-the-loop training for every processor to alleviate artifacts due to device mismatch and device non-linearity. Speeding up chip-in-the-loop training requires re-biasing the circuits in a manner that the analog functions remain invariant across training and inference. In this paper, we present an analog computational paradigm and circuits using "shape" functions that remain invariant to transistor biasing (weak, moderate, and strong inversion) and ambient temperature variation. We show that a core Shape-based Analog Compute (S-AC) circuit could be re-biased and reused to implement: (a) non-linear functions; (b) inner-product building blocks; and (c) a mixed-signal logarithmic memory, all of which are integral towards designing an ML inference processor. Measured results using a prototype fabricated in a 180nm standard CMOS process demonstrate bias invariance and hence the resulting analog designs can be scaled for power and speed like digital logic circuits. We also demonstrate a regression task using these CMOS building blocks.
READ FULL TEXT
