D2Computer Vision

Computer vision is the track that taught deep learning to see, and it remains the field where architectural ideas tend to appear first. The core problem is turning raw pixels into structured understanding: a label, a set of boxes, a per-pixel mask, or a full 3D scene. Convolution provides the inductive bias that makes this tractable, sharing weights across space so a network learns local patterns it can reuse everywhere.

You will build convolution and pooling from sliding inner products, then trace the architectural lineage from LeNet through ResNet's skip connections to ConvNeXt, and see how Vision Transformers reframe an image as a sequence of patches. From there you move into detection (R-CNN, YOLO, DETR), segmentation (U-Net, Mask R-CNN, Segment Anything), label-free pretraining with DINO and MAE, and neural rendering with NeRF and Gaussian Splatting.

These methods power everything from medical imaging to autonomous driving to the image encoders inside multimodal models, so the track connects directly to self-supervised learning and generative modeling later on.