Pseudoinverse Training for Neuromorphic Reservoir Computing Using Square-Root-Free Cholesky Decomposition

Loading...
Thumbnail Image

Files

TR Number

Date

2026-10-06

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Adaptive reservoir readouts require a compact training engine as well as an efficient inference circuit. This thesis develops a fixed-point, square-root-free readout trainer for a windowed echo state network used in MIMO-OFDM symbol detection. The trainer streams regularized sufficient statistics into a banked, packed memory, factors the 48-by-48 normal matrix in place with a reciprocal-based LDL-transpose decomposition, and reuses the factors for all four real outputs. A word-length analysis of the datapath shows that the statistics multipliers can be narrowed from 48 by 48 bits to 14 by 14 and 14 by 17 bits without changing any result bit, reducing DSP usage from 162 to 36 slices on the Zynq ZCU104. For the factorization, a standard eigenvalue argument shows that ridge regularization bounds every reciprocal operand by the inverse of the regularization constant. This bound sets the required operand range and explains why a smaller 32-bit factor/solve option, which uses 26 DSP slices, passes tests with strong regularization but rejects every recorded packet it was tested on at the regularization used for detection. The selected design, implemented in the detection configuration, uses 36 DSP slices, 31,488 LUTs, and 27 block-RAM tiles, meets a fully constrained 50 MHz out-of-context timing requirement with an estimated 69 mW of dynamic power, and completes a 540-sample training batch in 147,146 cycles, or 2.943 ms. In RTL simulation of 19,230 recorded frames from 30 recordings in five scenarios, it trains every frame and reproduces all 3,692,160 reference weights exactly. With these weights, the ESN detector reaches a bit error rate of 4.05 percent over all frames, against 5.78 percent for LS/ZF and 4.90 percent for baseline MMSE detection, and is the better of the three in four of the five scenarios; pilot-tracked conventional detectors remain more accurate. A triplet-STDP circuit study is also summarized, and the adaptation of the trainer to GlobalFoundries 40LP is identified as ongoing work.

Description

Keywords

Reservoir Computing, On-Chip Training, Square-Root-Free Cholesky Decomposition, Fixed-Point Hardware, MIMO Symbol Detection

Citation

Collections