Asked about Convolutional networks

Would you train a vision model from scratch?

An answer that works

Almost never. Start from a pretrained backbone, replace the head for your classes, and fine-tune. It needs far less data and compute and usually outperforms training from nothing unless your domain is very unlike the pretraining data.

Reading an answer is not being able to give one. The section that teaches this is Convolutional networks, 5 hours of Convolutional networks.

Learn this properly

Also asked about Convolutional networks

Why are convolutions suited to images?Have vision transformers replaced convolutional networks?What is the difference between detection and segmentation?How do you choose data augmentations?What is the output shape of a 3x3 convolution with stride 2 and padding 1 on a 64x64x32 input with 64 filters, and how many parameters does it have?How does CLIP enable zero-shot classification?Explain mean average precision for object detection.