Skip to content

🔧 Intermediate Tier: Deep Learning for Vision

Levels 3–5 · Goal: train, debug, and evaluate deep vision models · ~12–16 weeks at about an hour a day

This tier is being written

Pages marked "Coming soon" list exactly what each chapter will cover. Start with the Beginner tier.

You know how images work and how classical vision solves problems. Now you'll learn how deep networks learn to see, and master the three core tasks of modern vision: classification, detection, and segmentation.

What you'll be able to do

  • Explain what convolutional layers learn, and design and train CNNs that generalize.
  • Fine-tune pretrained models on small datasets of your own.
  • Visualize and debug what a network has learned, with Grad-CAM, saliency maps, and error analysis.
  • Build object detectors and evaluate them correctly with mAP.
  • Train segmentation models, and estimate pose and depth.
  • Use promptable foundation models like SAM, including as labeling tools.

The levels

  • Level 3: CNNs for vision

    From pixels to learned features, CNN architectures, training classifiers, transfer learning, and understanding CNNs.

    Start Level 3

  • Level 4: Object detection

    IoU, NMS, and mAP; two-stage and one-stage detectors; DETR; and detection in practice.

    Start Level 4

  • Level 5: Segmentation and dense prediction

    Semantic, instance, and panoptic segmentation; pose; depth; and promptable segmentation.

    Start Level 5

Tier checkpoint

You're ready for the Expert tier when you can do all of these without notes:

  • Compute a CNN's output shape, parameter count, and receptive field by hand.
  • Fine-tune a pretrained backbone, and explain your freezing and learning-rate choices.
  • Implement IoU and NMS, and explain how COCO mAP is computed.
  • Explain the difference between Faster R-CNN, RetinaNet, and DETR.
  • Train a U-Net, and report IoU and Dice correctly.

See the full roadmap for every concept in this tier.