Fig. 4: Stochastic gradient descent (SGD)-based neural network training based on bit-strings generated by TLQS.
From: Highly stable two-level current fluctuation in complex oxide heterostructures
a Simplified illustration of image super-resolution with VDSR. Note that, in this image super-resolution process, random number sequences are necessarily used for the random sampling of VDSR image dataset. b Schematics explaining the effect of random shuffling for SGD-based model training. The figure shows a two-dimensional parameter space depicting the loss landscape. The deeper (i.e., darker) the landscape, the lower the loss value. c, d The random shuffling enables faster and more accurate training. (c) Training loss and (d) validation accuracy curves for the network training with different shuffling methods. e–g The high-resolution images recovered from an identical low-resolution image with (e) no-shuffling, (f) Python’s NumPy shuffling, and (g) TLQS-based shuffling, respectively. The smooth and clear image recovered with the TLQS-based shuffling indicates the high quality of the random number sequences generated by the TLQS. Source data are provided as a Source Data file.
