Course Description
This course is a hands-on introduction to modern neural network models. We will cover the fundamentals of neural networks, and work our way up to modern generative architectures. We will cover stochastic gradient descent and backpropagation, along with related fitting techniques.
A significant component of this course will focus on interpretability: understanding how generative models manage to realize the functionality they offer. We will make use of the National Deep Inference Fabric (NDIF), a Northeastern-led project.
Logistics
I will assign homeworks to be done without AI, but because I cannot police this, they will not be graded. Instead, they are meant to help you learn the material and prep for the brief in-class assessments we will have regularly; these are intended to be straightforward if you have done the homeworks.Grading
| 35% | In-class quizzes and exercises |
| 25% | Midterm |
| 40% | Final project |
Your lowest quiz score will be dropped.
Prerequisites
Prior exposure to machine learning is recommended. Working knowledge of Python is required. Familiarity with linear algebra, basic calculus, and probability will be largely assumed, though we will review key prerequisites.
Midterm
An in-class midterm covering the foundations of deep learning.
Projects
A major component of the course is a project completed in pairs. Projects should concern some aspect of model interpretability and will use NDIF.
See project guidelines here
Academic Integrity and AI
No AI will be allowed while taking the in-class quizzes or midterm (all closed book, as well). You can use AI freely on homeworks (though these won't be graded and we recommend minimal use to ensure you understand the material) and on projects (but you must understand the code).
See the Northeastern Academic Integrity Policy for general academic integrity policy guidelines (which apply here).
Schedule
Dates and materials will be updated as the semester progresses. Readings reference d2l unless otherwise noted.
| Date | Topic | Readings | Notes | Materials |
|---|---|---|---|---|
| 9/9 (W) | Course aims, logistics; Review of supervised learning / Perceptron | d2l: Introduction; Original Perceptron (1957) | Join Piazza | Slides; Perceptron notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) |
| 9/14 (M) | Logistic Regression and Optimization via SGD | d2l: Preliminaries | Linear models notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) | |
| 9/16 (W) | Beyond Linear Models: The Multi-Layer Perceptron | d2l: MLPs (4.1) | Quiz 1; based on HW 1 | MLP notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude); Quiz 1 solutions |
| 9/21 (M) | Abstractions: Layers and Computation Graphs | d2l: Layers and blocks | Computation graphs notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) | |
| 9/23 (W) | Backpropagation I | d2l: Autodiff; Colah's blog; Rumelhart et al. (1986) | Backprop notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) | |
| 9/28 (M) | Backpropagation II | d2l: Backprop | Custom layer notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) | |
| 9/30 (W) | Optimizer matters: Training Neural Networks in Practice | d2l: Optimization | Quiz 2; based on HW 2 | Optimization notebook; Lecture notes: scribbled; Lecture notes: typeset (by Claude) |
| 10/5 (M) | Learning representations of discrete things: Embeddings | d2l: Word embeddings (14.1) | ||
| 10/7 (W) | Convolutional Neural Networks (CNNs) | d2l: CNNs (6.1) | ||
| 10/12 (M) | No class — Indigenous Peoples Day | |||
| 10/14 (W) | Stacking ConvNets, residual connections, and other tricks | d2l: CNNs (6.2–6.5); Modern CNNs (7.1, 7.5–7.7) | Quiz 3; based on HW 3 | |
| 10/19 (M) | Recurrent Neural Networks I | d2l: RNNs (8.1, 8.4); Karpathy: Unreasonable Effectiveness of RNNs | ||
| 10/21 (W) | Recurrent Neural Networks II | d2l: RNNs (8.7) | ||
| 10/26 (M) | Transformers, self-supervision, contextualized embeddings | d2l, Ch. 11 | ||
| 10/28 (W) | More Transformers; BERT and BERTology | Quiz 4; based on HW 4 | ||
| 11/2 (M) | Midterm review | |||
| 11/4 (W) | Midterm | |||
| 11/9 (M) | Posttraining (instruction tuning and RLHF) 1 | |||
| 11/11 (W) | No class — Veterans Day | |||
| 11/16 (M) | Posttraining (instruction tuning and RLHF) 2 | Project proposals due | ||
| 11/18 (W) | How do LLMs work? Interpretability | |||
| 11/23 (M) | Guest lecture: Activation verbalization | |||
| 11/25 (W) | No class — Fall break | |||
| 11/30 (M) | Diffusion Models | Step-by-Step Diffusion: An Elementary Tutorial | Quiz 5; based on HW 5 | |
| 12/2 (W) | Ethical problems with generative models | |||
| 12/7 (M) | Dedicated project feedback and help | |||
| 12/9 (W) | Project presentations |