Unit 10 of 15

Representations, Vision, and Learning from Few Labels

  • 5 lessons
  • Videos 67 to 71

Learned representations let a model reuse what was found in another task or in unlabeled data. This unit covers convolutional networks, transfer learning, autoencoders and learning with few labels.

A convolution slides one small filter across an image, and stride, padding and pooling decide the output shape and receptive field. Embeddings are learned feature vectors, and transfer learning freezes a pretrained feature extractor or fine-tunes it. Autoencoders compress and reconstruct inputs, and a linear one with squared loss can recover the principal component subspace. Pseudo-labeling, consistency regularization and contrastive learning use unlabeled data, which can hurt when assumptions fail, and active learning chooses examples to label. An image classification workflow ends the unit.

Every lesson will be one video. The videos are coming soon, and each lesson page already has the definitions and an example for every key term.

Convolution and Transfer Learning

  1. Convolutional Neural Networks

    How a convolutional layer slides one small filter across an image to build a feature map, how stride and padding set the output size, how receptive fields grow with depth, and what pooling keeps and discards.

    Video 67

  2. Embeddings and Transfer Learning

    What an embedding is and how to compare two of them, and how transfer learning reuses a pretrained model as a frozen feature extractor with a new prediction head or adapts it by fine-tuning.

    Video 68

Learning from Unlabeled Data

  1. Autoencoders and Nonlinear Representations

    How an autoencoder squeezes each input through a bottleneck and rebuilds it, how reconstruction loss is reported, how a linear autoencoder relates to PCA, and what denoising and nonlinear reduction add.

    Video 69

  2. Semi-Supervised and Self-Supervised Learning

    How semi-supervised learning uses unlabeled inputs through pseudo-labeling and consistency regularization, how self-supervised and contrastive learning construct their own targets, and how active learning picks labels.

    Video 70

Image Classification in Practice

  1. Image Classification and Augmentation

    How to build and evaluate a small image classifier with label-preserving data augmentation, a subject-level split, a simple baseline and a frozen-feature model, and how far its test score reaches.

    Video 71