Fig. 5

Micro-architectural improvements to the MobileNetV3 block improve training dynamics and reduce memory requirements at minimal cost to inference speed. Neural networks were trained for 6000 parameter updates for each dataset and micro-architecture. A Loss curves for the touch-evoked response dataset. B Loss curves for the visuomotor response dataset. C Training speeds. D Inference speeds. E Parameter counts. F Theoretical minimum memory footprints. Data represent the mean or the mean plus and minus the standard error of the mean of 10 independent experiments in which datasets were partitioned and neural network weights were initialized using different random seeds