Reuse one small test everywhere

A kernel is a small learned filter. Choose its width, then move the same filter across every time position.

Kernel width Jump size (stride) Edge handling (padding)

6 answer boxes (outputs)

input x · x₁ means the first value · orange badges show the kernel weights

answer track y · y₁ means the first score · choose one to move the window

Multiple channels: one kernel makes one output track. Many kernels make parallel tracks that can detect different sound patterns.

Model stages (layers): 1

CNN means convolutional neural network: a model that reuses learned local filters. Width 1 transforms one position; width 3 combines three neighbors. Translation equivariance means that, with stride 1 and away from the edges, shifting a pattern shifts its filter response. Padding adds edge zeros, and stride is the filter’s jump size.

An interactive sliding three-number pattern test over eight made-up sound values.

Reuse one small test everywhere