A ground-up tour of convolutions and how they compose into CNNs. Uses fastai to cut boilerplate so the focus stays on the concepts.

2D convolution sliding window

Key ideas covered:

  • The convolution operation: sliding kernels, dot products, feature maps
  • Filters as feature detectors: edges, textures, and learned representations
  • Padding & stride: controlling output dimensions and downsampling
  • Pooling: spatial invariance and dimensionality reduction
  • Receptive fields: how deep layers see larger regions of the input
  • Training a CNN: end-to-end with fastai’s training loop