CS342: Neural Networks

In this course, we discuss the basics of deep neural networks. We will look at different neural architectures, as well as how to train, tune, and test networks. We will cover both the theory and practice of deep learning, using hands-on implementations in PyTorch. We will then briefly look at a number of common applications of deep learning including computer vision, sequence modeling, deep reinforcement learning, and generative modeling. Over the course of the homework assignments, we will develop a vision system for a racing simulator, SuperTuxKart.

Logistics

The class meets Monday and Wednesday from 2:00 to 3:30. We will be meeting on Zoom for the foreseeable future (see Canvas for Zoom links), but if we are able to come back to in-person learning, then we will meet in JGB 2.216.

We will use Piazza for discussions and Canvas for homework. We will start each Monday class with a quiz, and then we will go through the quiz questions in class. After that, we’ll generally use about half the class time for lecture and half to start on a coding exercise related to the topics we just covered. The coding exercise will be done either in small groups or as a class discussion. On Wednesdays we will again use approximately half the class for lectures and half to continue the coding exercise from Monday.

Calendar

This calendar should be considered tentative. While we plan to stick to this schedule as closely as possible, we may delay or reorder topics as issues come up during the semester. The homework due dates will not be moved earlier.

Date Topic Exercise Resources
1/19 Introduction, background Slides
1/24 Background - Tensors, broadcasting KNN Slides, Notebook
1/26 Background - Probability KNN Slides
1/31 Models - Linear, computation graphs Linear classifier Slides, Notebook
2/2 Training - Optimization, loss Linear classifier Slides
2/7 Models - Nonlinearities MLP Slides
2/9 Models - Layers, activation, hyperparameters MLP Slides, Supplement
2/13 HW1 is due at 11:59 PM Assignment
2/14 CNN - Convolution, pooling CNN Slides
2/16 CNN - Receptive field, principles CNN Slides
2/21 CNN - Dilation, upconvolution, ResNets CNN Blocks Slides, Supplement
2/23 Practicalities - Data splitting, Initialization CNN Blocks Slides, Supplement, Notebook
2/27 HW2 is due at 11:59 PM Assignment
2/28 Pracitcalities - Normalization Image gen. Slides, Notebook
3/2 Practicalities - Overfitting Image gen. Slides, Notebook
3/7 Practicalities - Overfitting Binary segment. Slides
3/9 Practicalities - Learning rate, Optim Algorithms Binary segment. Slides, Notebook
3/14 Spring break
3/16 Spring break
3/20 HW3 is due at 11:59 PM Assignment
3/21 Vision - Classification Keypoint est. Slides
3/23 Vision - Object detection Keypoint est. Slides
3/28 Vision - Segmentation Keypoint est. Slides
3/30 Vision - Temporal models Keypoint est. Slides
4/3 HW4 is due at 11:59 PM Assignment
4/4 RL - Imitation learning Imitation Slides, Notebook
4/6 RL - Policy gradients, gradient free optim. Imitation Slides, Notebook
4/11 Sequence - RNN, LSTM, GRU Imitation 2 Slides
4/13 Sequence - Temporal conv., WaveNet Imitation 2 Slides, Notebook
4/17 HW5 is due at 11:59 PM Assignment
4/18 Special topics - Bias, Fairness, and Ethics RL Slides
4/20 Special topics - Game Playing RL Slides
4/25 Special topics - Generative Models Slides
4/27 Special topics - Safe RL Slides
5/2 Final presentations
5/4 Final presentations
5/8 Final project, makeup homework due at 11:59 Make-up homework, Final project

Grading

  • Quizzes: 10% (the 2 lowest-scoring quizzes will be dropped).
  • Coding/Class participation: 10% (you may miss 3 classes without an excuse).
  • Homework: 50% (10% per assignment with the lowest score being dropped).
  • Final project: 30%

Late Policy

Quizzes cannot be made up if missed since we discuss the solutions immediately. Homework may be turned in one day late at a 25% penalty or two days late for a 50% penalty. In addition, you will have three slip days for homeworks which will be applied automatically. (For example, the first time a homework is one day late, you will be graded as if that homework were turned in on time, and you will lose one slip day. You would then have two slip days remaining for the rest of the semester.) After three days homeworks will not be accepted because the homework solutions will be released. The final project cannot be turned in late.

Academic Honesty

Homeworks and quizzes are to be done individually. For the final project, groups of up to four will be allowed. You are allowed to discuss ideas and share data, but you may not share any code. You also may not post solutions to any assignments in a place where other people may see them. You are allowed to look at publically available online discussions (e.g., StackOverflow) so long as you cite any code or ideas you take from such discussions. If you are unsure about whether a particular behavior would consistute academic dishonesty, don’t hesitate to ask the instructor. For more information on UT’s academic honesty policy, please refer to the departmental guidelines.

Documented Disability Statement

The University of Texas at Austin guarantees that students with disabilities have access to appropriate accommodations. You may request an accommodation letter from the Division of Diversity and Community Engagement, Services for Students with Disabilities https://diversity.utexas.edu/disability/.

If you have approved accommodations for the course, please contact us to arrange them. Please do this as soon as possible, so that you can have the benefit of the accommodations throughout the duration of the course.

Behavioral Concerns

If you are worried about someone who is acting differently, you may use the Behavior Concerns Advice Line to discuss by phone your concerns about another individuals behavior. This service is provided through a partnership among the Office of the Dean of Students, the Counseling and Mental Health Center (CMHC), the Employee Assistance Program (EAP), and The University of Texas Police Department (UTPD). Call 512-232-5050 or visit https://besafe.utexas.edu/behavior-concerns-advice-line

Acknowledgement

Note that most of this syllabus (with the exception of the calendar, the logistics, and some wording changes) was developed by Philipp Krähenbühl and is reused here with permission.