Introductory Lecture Notes on
Deep Learning
Abstract
These lecture notes are for a one-semester (12-week) introductory course on deep learning, covering both theory and practice. Students will learn the foundational mathematics behind deep learning and explore topics such as multi-layer perceptrons (MLPs), back-propagation and automatic differentiation, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformers. These techniques play a crucial role in modern artificial intelligence (AI) systems, including image and video understanding, natural language processing, generative AI, robotics, medicine and scientific discovery. The course includes various practical assessments to enhance student’s understanding and intuition of deep learning and its diverse applications. Students are expected to have strong programming skills and previous exposure to linear algebra, differential calculus, and probability theory. Assessment details do not form part of these notes.
Appetizer
Consider the two-dimensional shape depicted in Figure 1. Nobody would have trouble recognising the shape as a triangle despite the lines not being completely straight and them not meeting precisely at each corner. And when asked why the shape is a triangle, most people would answer that it’s because the shape has three sides. Deep learning, however, takes a different perspective. Deep learning says that the figure is a triangle because it looks like a triangle. In the same way, a cat is a cat because it looks like a cat, and a dog is a dog because it looks like a dog. No further justification is needed.2 This is because deep learning methods learn from data, i.e., from training examples, rather than from prescribed rules. Cognitive scientists call this reasoning by analogy as opposed to reasoning by deduction (rules). In this course we will study deep learning and discover how such reasoning is at all possible for a machine.
Contents
- Introduction
- Linear Classification and Multilayer Perceptrons
- Back-propagation and Learning
- Convolutional Neural Networks
- Object Detection and Image Segmentation
- Sequence Models
- Developing, Debugging and Diagnosing
- Transformers
- Contrastive Learning
- Deep Reinforcement Learning
- Generative AI: VAEs, Diffusion and Flows
- 1. We give warm thanks to the many students and colleagues who have provided feedback and corrections on earlier versions of these notes. we are especially grateful to Yicong Hong for creating this navigable website version of the notes.
- 2. This is not completely true—there is an entire research field devoted to the important topic of explainable AI, which requires that a human understandable justification be given for the outputs, answers, predictions, and decisions produced by an AI algorithm.