Chapter 4 · The deep learning era (2006–2016)
Convolution
A convolution slides a small filter over an image to detect local patterns like edges.
Chapter 4 · The deep learning era (2006–2016)
A convolution slides a small filter over an image to detect local patterns like edges.
The amber square is the 3×3 window. At each spot it multiplies the pixels under it by the filter and adds them up. Where the picture changes from dark to bright, the answer is big.
Each pixel is a number: 0 for dark, 1 for bright. A computer sees only this grid.
Drag on the left grid to draw or erase. Choose a filter. Then move the spot slider to see the nine multiplications behind any single output pixel.
Lights up where dark meets bright going left to right.
1×-1 + 1×0 + 1×1 + 1×-2 + 1×0 + 1×2 + 1×-1 + 1×0 + 1×1 = 0
A teaching version: a 10×10 black-and-white picture and zeros beyond the border. Violet shows positive answers and coral negative ones. Real networks learn their filters instead of using fixed ones.
What does a convolution do at each spot?
A vertical-edge filter looks at a perfectly flat patch. The answer is…
Why is one small filter enough for a whole picture?