Hands-on projects that turn theory into skill — each with a dataset, architecture, and a concrete build plan.
Build and train a CNN on CIFAR-10, then compare against a fine-tuned ResNet.
Annotate a small dataset and fine-tune YOLO to detect your own classes.
Train a U-Net to segment lesions or organs and evaluate with Dice/IoU.
Implement a ViT end-to-end — patch embeddings, attention, and training.
Train a DDPM to generate images and watch samples emerge from noise.
Combine Grounding DINO + SAM to segment anything from a text prompt.
Estimate depth from single images and build a 3D point cloud you can rotate.
Search a photo library with natural language using CLIP embeddings.
Detect and track people with keypoints from a webcam in real time.