Fig. 3: Parallelization of HGE in a RNPU.
From: Gradient descent in materia through homodyne gradient extraction
a Schematic representation of the RNPU (yellow dots represent boron atoms) and the two different ways to extract the gradient of the output current with respect to the input voltages using HGE: single-input and multi-input. The rightmost panel shows typical experimental output currents for both cases. b Analytical description of expected mean squared error vs. frequency spacing Δf. Horizontal dashed line: mean squared error for the sequentially extracted derivative. Solid black line: mean squared error for the derivative extracted in parallel, for a channel with an output signal strength that is 2.05 times weaker than its neighbouring frequencies. Upper limit of grey area: the same, but for a channel with an output signal strength that is 20 times weaker than the others. Lower limit of grey area: the same, but for a channel with an output signal strength that is as strong as all the others. The magnitude of the noise is fixed to a value of 10−5 nA2 Hz−1 at 1 kHz (see Fig. 2c). Side panels: Schematic power spectral density (PSD) graphs of a parallel mixed HGE signal, demonstrating how the frequency spacing Δf affects the final filtered signal. c, d Comparison between the single- and multi-input HGE derivatives for 1000 sets of random input voltages applied to the RNPU with a frequency spacing of either 10 Hz (c), or 40 Hz (d). In both cases the black line indicates y = x. The measurement time per data point is 0.1 s both in the calculations and the measurements.
