Deep Learning · TEE · guide

Convolutional neural networks

  1. Reading image shapes before doing CNN arithmetic
  2. A computer sees an image as numbers
  3. The convolution calculation
  4. Stride padding and floor
  5. Convolution over several channels
  6. One filter over multiple channels: all contributions go into one cell
  7. Choosing more or fewer filters
  8. Pooling needs comparisons or averages
  9. Parameters are different from activation values
  10. One complete CNN pipeline
  11. CNN versus a fully connected image network
  12. Model answer: CNN structure and parameter sharing
  13. Output size and parameter count are different
  14. Model answer: shape, activations and parameters
  15. Faculty example: a complete binary-image convolution
  16. Faculty example: combine three RGB channel sums
  17. Max, min and average pooling on the same lecture grid
  18. Faculty CNN table: shapes, activations and 62958 parameters
  19. Even-sized kernels: SAME padding is not always symmetric