Vision and audio

Convolution and transfer learning, and the lesson that a pretrained model beats training from zero.

An image as a tensorChannels, dimensions and normalisation, and how a pixel becomes a number.
ConvolutionFilter, stride and padding, and what the first layer actually sees.
Transfer learningFreezing layers and swapping the head, the lesson that saves you weeks.
Data augmentationWhat helps, what destroys meaning, and why flipping Arabic text is a disaster.
Vision transformersAn image as a sequence of patches, and when ViT beats convolution.
AudioMel spectrograms, speech recognition, and where Arabic support falls short.
Project: an image classifierClasses you collect yourself, two versions, and a time and accuracy comparison.