CS 7150: Deep Learning - Fall 2024

Time & Location:

12:30 - 2:10pm Eastern Time, Tuesdays, along with additional asynchronous work.
Location: See Canvas for Zoom link.

Staff

Instructor: Paul Hand
Email: p.hand@northeastern.edu     Office Hours: Fridays 4-5 PM Eastern. See Canvas for Zoom link.

TA: Daniel Golfarb
Email: goldfarb.d@northeastern.edu     Office Hours: Thursdays 2-3 PM Eastern. See Canvas for Zoom link.

TA: Sean Gunn
Email: gunn.s@northeastern.edu     Office Hours: Mondays 3-4 PM Eastern. Location: WVH 102 (bring your NU ID), or optionally see Canvas for Zoom link.

Course Description

Note: This differs from the official course description. Please read carefully.

Introduction to deep learning, including the statistical learning framework, empirical risk minimization, loss function selection, fully connected layers, convolutional layers, pooling layers, batch normalization, multi-layer perceptrons, convolutional neural networks, autoencoders, U-nets, residual networks, gradient descent, stochastic gradient descent, backpropagation, autograd, visualization of neural network features, robustness and adversarial examples, interpretability, continual learning, and applications in computer vision and natural language processing. Assumes students already have a basic knowledge of machine learning, optimization, linear algebra, and statistics.

Overview

The learning objectives of this course are that students should:

• be able to train neural networks at a variety of tasks
• be able to select appropriate neural network architectures for a variety of tasks
• have read and be able to discuss the contents and contributions of important papers in the field
• have a basic theoretical understanding of the tools of deep learning

Course Structure and Expectations:

This course will have two components: a synchronous meeting over Zoom on Tuesdays and asynchronous discussion throughout the rest of the week. During the synchronous session, the instructor will be screensharing papers and/or digital notes from the classroom and will facilitate classwide discussion over Zoom. Participation in the class discussion is expected, and everyone will have regular turns at speaking. During class, students should be fully engaged with the class, to the best of their ability. In particular, students should be willing to share with the whole class their responses to prepared questions. Students are encouraged to leave their video connection on for the majority of class and should mute themselves when they are not speaking. Lectures will be recorded and posted on Canvas.

Student Work:

Students will be expected to complete the following work:

• Each week, students will be given a set of questions to answer in writing about that week's content. Students are expected to submit the answers to these questions to Gradescope by 12:30 PM on the relevant day of class in order to ensure they are prepared to engage in the class discussion. Students will work in small groups.
• Students are expected to participate in the whole-class discussions by presenting prepared responses and answering spontaneous questions.
• Each week, the TAs will grade student responses to two of the preparation questions. Those questions will be announced in class, and students are able to revise their answers based on the class discussion until Friday at 9 PM Eastern.
• There will be three homework assignments throughout the semester. They will involve both pencil-and-paper work and computation involving training neural networks. The results will be submitted to gradescope in the form of a pdf.
• Students will complete a project. They will reproduce an empirical observation in a paper of their choice by independently recreating and executing code to train and test that network. You may not use the code from the specific paper you are replicating, but you may use code from other papers.

Course grades:

Course grades will be based on: 30% Preparation questions for classroom, 20% Participation, 30% HWs, 20% Project.
Letter grades will be assigned on the following scale: 93%+ A, 90-92% A-, 87-89 B+, 83-87% B, 80-82% B-, 77-79 C+, 73-77% C, 70-72% C-,60-70% D, 0-59% F.

Prerequisites

Students are expected to have experience with a class in Machine Learning. The class will assume students are comfortable with some linear algebra, probability, and statistics. Some experience with neural networks, python, PyTorch/TensorFlow will be helpful but can be acquired while taking this class. If you do not have experience with PyTorch, a good resource is the book Deep Learning with Pytorch.

Day	Date	Class Discussion Will Be On:
1	T 9/10	Machine Learning Review (Notes) A DARPA Perspective on Artificial Intelligence Preparation Questions for Class Class Notes
2	Asynch	Machine Learning Review (Notes)
3	T 9/17	Deep Learning Understanding deep learning requires rethinking generalization Preparation Questions for Class (pdf, tex)
4	Asynch	Architectural Elements of Neural Networks. (Notes).
5	T 9/24	Visualizing and Understanding Convolutional Networks Visualizing Higher-Layer Features of a Deep Network Preparation Questions for Class (pdf, tex)
6	Asynch	Gradient Descent and Stochastic Gradient Descent (Notes)
7	T 10/1 HW 1 (pdf, tex) DUE FRI 10/4	Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift How Does Batch Normalization Help Optimization? Preparation Questions for Class (pdf, tex)
8	T 10/8	Deep Learning Book - Chapter 8 Adam: A Method for Stochastic Optimization Preparation Questions for Class (pdf, tex)
9	Asynch	Neural Network Architectures for Images (Notes)
10	T 10/15	Deep Residual Learning for Image Recognition (ResNets) ImageNet Classification with Deep Convolutional Neural Networks (AlexNet) Preparation Questions for Class (pdf, tex)
11	Asynch	Watch DeepLearningAI videos on Object Localization and Detection: 1, 2, 3, 4, 6, 7, 8, 9, 10, You Only Look Once: Unified, Real-Time Object Detection
12	Asynch	Adversarial Examples for Deep Neural Networks (Notes)
13	Asynch HW 2 (pdf, tex) DUE (F 10/25)	Work on HW
14	T 10/29	Explaining and Harnessing Adversarial Examples Robust Physical-World Attacks on Deep Learning Models Preparation Questions for Class (pdf, tex)
15	Asynch	Continual Learning and Catastrophic Forgetting. (Notes).
16	T 11/5	Overcoming catastrophic forgetting in neural networks Preparation Questions for Class (pdf, tex)
17	Asynch	YouTube Video: LSTM is dead. Long Live Transformers. Watch the first 23 minutes. YouTube Video: Illustrated Guide to Transformers Neural Network: A step by step explanation
18	T 11/12 HW 3 (pdf, tex) DUE F 11/15	Attention Is All You Need Language Models are Few-Shot Learners Preparation questions for class (pdf, tex)
19	Asynch Project Planning Document (pdf, tex) Due (M 11/18)	Automatic Differentiation, Backpropagation Watch this video of automatic differentiation. Watch this lecture (from start until time 38:30) on backpropagation of neural networks.
20	T 11/19	Generative Adversarial Networks (notes)
21	Asynch	Variational Autoencoders (notes)
22	T 11/26	Blog post on Score Based Diffusion Models
23	T 12/3	Project Presentations