class: center, middle # Course Overview Edouard Yvinec .affiliations[     ] --- We saw: - the basic logisitic regression -- - auto-grad and automatic differentiation -- - conv nets -- - classification, object detection, and image segmentation -- - GAN, VAE, normalizing flows and diffusion models -- - transformers and mlp mixers -- - tools to adapt such models to videos --- # When should you use ConvNets? -- - Convolutional neural networks are very good for images as they are translation invariant. - CNNs are very effective (few parameters and strong expressivity) - CNNs can be massively parallelized - CNNs are good at capturing biases in the data --- # How to select the task given a real problem? What should you use ? object detection, semantic segmentation, classification, regression? -- - Always a trade-off between having a lot of data and having a good quality annotation - Check if there are standard dataset with a close distribution to your target use-case - The finer the predictions the less likely you are to need to upgrade your dataset and annotation (future proof) - Sometimes, you don't have to choose... - Prioritize end-to-end processes when possible for accuracy - Prioritize sequential processes for explainability --- # When should you use generative models? -- - data augmentation - GANs can be used to learn to be ivnariant to some information (eliminate unwanted biases) - super-resolution, when your recorder has a poor quality - test the limit of your model (adversarial attacks) --- # How can you improve a baseline? -- - hyper-parameter fitting - modern training protocols (large batches, mix-up, Nadam or AdamW optimizers,...) - self-attention layers - adequate tricks for the target task (learn from a downstream task, use physics to solve the problem,...) - pre-training (larger dataset, may be unlabelled) --- # Real-case scenario ### Count cars on a road -- What are strong baselines? -- How do we test them? -- How do you propose to build your own baseline? -- How do we test it? -- How can we improve it? --- # Real-case scenario \#2 ### Detect defects on a production line -- What are strong baselines? -- How do we test them? -- How do you propose to build your own baseline? -- How do we test it? -- How can we improve it?