A two-layer network's behavior as width grows depends on a single scaling factor in front of its output. I compared the three standard choices, Neural Tangent Kernel, Mean Field, and Random Features, across three datasets (synthetic, MNIST, CIFAR-10) and a range of widths, activations, and learning rates.

• Mean Field reached the lowest test loss across the board, consistent with its learning rate scaling with width.
• Random Features, which only trains the outer weights, performed worst, as expected from learning fewer features.
• NTK landed in between, with output variance that stays stable rather than shrinking to zero as width grows.

A writing-course paper comparing three signals used to detect movement intent for exoskeleton control in neurorehabilitation: EEG, surface EMG, and a fusion of the two. I compared accuracy, speed, and equipment cost across studies using each one.

• sEMG won on all three (96% accuracy, 59.9ms, ~US$672 in sensors), since most of EEG's cost comes from cleaning up its noise.
• EEG + sEMG fusion is slower and pricier, but more robust early in rehabilitation when a patient's sEMG signal is still weak.
• The studies compared aren't directly comparable (different subject counts, classes, equipment), so the conclusions are more suggestive than definitive.