Deep learning arrives for images. CNNs replace densely connected layers with learned spatial filters — convolutions that detect edges, textures, and shapes regardless of where they appear in the image. This unit unfolds over several weeks and is also when the spring research project gets seeded: you’ll pitch an idea, refine it, and present progress.

Concepts You’ll Learn About

  • Convolution — sliding a filter across an image; detecting local patterns; feature maps
  • Pooling — downsampling feature maps; translation invariance
  • CNN architecture — conv → activation → pool → flatten → dense; stacking layers to build hierarchy
  • Data augmentation — flipping, cropping, and rotating training images to reduce overfitting
  • Transfer learning (first look) — how networks trained on one dataset can be repurposed; previews Unit 11
  • AlexNet — the 2012 architecture that started the deep learning era; depth, ReLU, and dropout

Topics

Convolutional neural networks are specifically designed to operate on 2D data. Before introducing this new model, let’s first walk through a complete example of an image-processing workflow using the dense neural network layers you already know. The CIFAR-10 dataset contains thousands of color images labeled with one of 10 categories. Work through to see how to load, train and improve the model using dense layers.

As a follow-up assignment, find a TensorFlow Datasets classification task and replicate the full pipeline in the example CIFAR notebook.

Next explore what an image convolution is

  • LeNet — the foundational CNN architecture applied to MNIST/FashionMNIST. Following d2l.ai chapter 7.6 (local PDF). A great 3D interactive visualization accompanies this session.

  • Activation functions — covered during Engineering Week; see Unit 09 for the notebook.

  • AlexNet on CIFAR-100 — scaling depth and applying dropout; training a historically significant architecture on a harder dataset.

  • Presentations — second round.

What’s next

Unit 11 introduces transfer learning — the insight that you don’t have to train from scratch — and ends with a pivot toward sequential data.