Teaching assistants
Pingchu Zhang and Chenxin Lyu
In-person and online office hours rotate weekly. Open the schedule below for dates, times, and locations.
Fall 2026
A graduate-level course on the foundations, architectures, optimization methods, and modern applications of deep learning.
Course at a glance
This page is the public home for the course topics, slide releases, and assignments. Enrollment-specific information, submissions, deadlines, and grades are maintained in Waterloo Learn.
Course support
Teaching assistant office hours
Go to the TAs for: questions about assignments or coding.
Teaching assistants
In-person and online office hours rotate weekly. Open the schedule below for dates, times, and locations.
| Week starting | In-person session: date, time, TA and room | Online session: date, time and TA |
|---|---|---|
| Oct. 5 |
Tuesday, Oct. 6 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Oct. 8 9–10 a.m. · Chenxin |
| Oct. 19 |
Tuesday, Oct. 20 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Oct. 22 9–10 a.m. · Chenxin |
| Oct. 26 |
Monday, Oct. 26 9–10 a.m. · Chenxin · M3 4221 |
Thursday, Oct. 29 9–10 a.m. · Pingchu |
| Nov. 2 |
Tuesday, Nov. 3 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Nov. 5 9–10 a.m. · Chenxin |
| Nov. 9 |
Monday, Nov. 9 9–10 a.m. · Chenxin · M3 4221 |
Thursday, Nov. 12 9–10 a.m. · Pingchu |
| Nov. 16 |
Tuesday, Nov. 17 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Nov. 19 9–10 a.m. · Chenxin |
| Nov. 23 |
Monday, Nov. 23 9–10 a.m. · Chenxin · M3 4221 |
Thursday, Nov. 26 9–10 a.m. · Pingchu |
| Nov. 30 |
Tuesday, Dec. 1 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Dec. 3 9–10 a.m. · Chenxin |
| Dec. 7 |
Monday, Dec. 7 9–10 a.m. · Chenxin · M3 4221 |
Thursday, Dec. 10 9–10 a.m. · Pingchu |
| Dec. 14 |
Tuesday, Dec. 15 10–11 a.m. · Pingchu · Desk #3, M3 2101 |
Thursday, Dec. 17 9–10 a.m. · Chenxin |
| Dec. 21 |
Monday, Dec. 21 9–10 a.m. · Chenxin · M3 4221 |
Thursday, Dec. 24 9–10 a.m. · Pingchu |
Online office hours are accessed through the Zoom widget in LEARN.
From the course outline
Required textbook
Benyamin Ghojogh and Ali Ghodsi. Springer, 2026. ISBN 978-3-032-10738-1 (eBook).
View textbook| Component | Weight |
|---|---|
| Assignments | 45% |
| Paper presentation | 10% |
| Group final project | 45% |
| Total | 100% |
| Assessment | Release | Due |
|---|---|---|
| Assignment 1 | Monday, October 5 | Monday, October 26, 11:59 p.m. |
| Assignment 2 | Monday, October 26 | Monday, November 16, 11:59 p.m. |
| Assignment 3 | Monday, November 16 | Monday, November 30, 11:59 p.m. |
| Paper presentation | Eligible venues specified in the course outline | Scheduled presentation date |
| Group final project | To be announced | Monday, December 21, 11:59 p.m. |
All dates are tentative and may change. Changes will be announced.
Latest updates
Course site
This page will be updated throughout the term. Slides and assignments will be posted below as they are released.
Communication
Questions about lectures and course content should be posted on Piazza so that answers are available to the class.
Course schedule
This tentative schedule follows the lecture sequence in the course materials; textbook references provide supporting readings.
| Lecture content | Textbook alignment | Materials |
|---|---|---|
| Lecture 1 - Motivation, history, machine-learning foundations, and a neural-network overview | Ch. 1, Secs. 1-5; Ch. 2, Secs. 2.1-2.7 | Lecture 1 slides |
| Lecture 2 - Neural-network history, McCulloch-Pitts networks, the perceptron and linear separability, feedforward neural networks, and backpropagation | Ch. 1, Sec. 5; Ch. 2, Secs. 2.2-2.7; Ch. 3, Secs. 2-3 | Lecture 2 slides |
| Lecture 3 - Stochastic gradient descent, mini-batches, learning-rate tuning, momentum, optimizer selection, and first- versus second-order methods | Ch. 3, Secs. 2-5; the modern optimizer survey is supplemental | Lecture 3 slides |
| Lecture 4 - Model selection, empirical and true error, SURE, double descent, implicit regularization, and grokking | Ch. 5, Secs. 2-5; Ch. 6, Sec. 5; optional SURE proof: Ch. 5, Sec. 13 | Lecture 4 slides |
| Lecture 5 - Regularization: weight decay, augmentation and noise injection, manifold tangent classifier, early stopping, parameter sharing, label smoothing, bagging, and dropout | Ch. 5, Secs. 6-10.3; manifold tangent classifier, parameter tying/sharing, and general label smoothing are supplemental | Lecture 5 slides |
| Lecture 6 - Dropout for linear regression, batch normalization, optimization effects, and normalization alternatives | Ch. 5, Secs. 10.3-11; normalization alternatives and modern BatchNorm optimization analysis are supplemental | Lecture 6 slides |
| Lecture 7 - Convolution and cross-correlation, sparse interactions, parameter sharing, equivariance, pooling, CNN training, ResNet, and DenseNet | Ch. 4, Secs. 1, 3.1.1-3.1.4, 3.2.1, 3.2.3-3.2.7, 4-4.2, and 5.6-5.7 | Lecture 7 slides |
| Lecture 8 - RNNs and parameter sharing, backpropagation through time, vanishing/exploding gradients, echo-state networks, long delays and leaky units, LSTM, GRU, and gradient clipping | Ch. 8, Secs. 1-3.4, 4.1, and 4.3.1 | Lecture 8 slides |
| Lecture 9 - Sequence-to-sequence models, attention for machine translation, self-attention, queries/keys/values, attention scoring and matrix formulation, and image captioning | Ch. 9, Secs. 1-2.5 | Lecture 9 slides |
| Lecture 10 - Transformer encoder/decoder, multi-head attention, position-wise feedforward networks, residual connections and layer normalization, masked self-attention, cross-attention, word prediction, and sinusoidal positional encoding | Ch. 9, Secs. 3.1-3.3.3; attention review: Secs. 2.4.2-2.4.3 | Lecture 10 slides |
| Lecture 11 - A unified view of MLPs, CNNs, attention, RNNs, and state-space models; linear attention and KAN; BERT pretraining, fine-tuning, and sentence embeddings; GPT, T5, and major language-model families | Ch. 9, Secs. 2.4 and 4.2.1; Ch. 10, Sec. 7; Ch. 11, Secs. 2, 3.1-3.2, 3.4, and 4-5. Unified framework, KAN, and newer model updates: lecture slides. | Lecture 11 slides |
| Lecture 12 - LLM alignment and InstructGPT; supervised fine-tuning, reward modeling, reinforcement-learning foundations and policy gradients, RLHF with KL regularization, and direct preference optimization (DPO) | Ch. 11, Secs. 7-9; reinforcement-learning foundations and policy gradients: Ch. 18, Secs. 2 and 9 | Lecture 12 slides |
| Lecture 13 - Performer attention, random-feature kernel approximations and FAVOR+, normalization and linear-time computation; variational inference, KL divergence and ELBO derivation, and a brief VAE introduction | Ch. 9, Sec. 4.2; Ch. 12, Secs. 1-2.1 | Lecture 13 slides |
| Lecture 14 - GAN training and objective functions, optimal discriminator and Jensen-Shannon divergence, mode collapse and minibatch discrimination, conditional GANs, adversarial autoencoders, and image-generation applications | Ch. 14, Secs. 2.1-2.4, 2.6-2.7, 3.1-3.3, 7, and 8.1 | Lecture 14 slides |
| Lecture 15 - Diffusion models and DDPMs; forward noising and reverse generation, Gaussian posterior and noise prediction, training and sampling, VAE comparison, and ELBO decomposition | Ch. 15, Secs. 1 and 2.1-2.7 | Lecture 15 slides |
| Lecture 16 - Graph data and graph-, node-, and edge-level tasks; graph Laplacians and spectral graph cuts, graph Fourier transforms, spectral convolution, and ChebNet with Chebyshev polynomial filters and local recurrence | Ch. 17, Secs. 1-3 | Lecture 16 slides |
| Lecture 17 - GCNs, neighborhood aggregation, graph learning tasks, GATs, and the Transformer connection | Ch. 17, Secs. 3-6.2 | Coming soon |
| Lecture 18 - PAC/VC generalization theory, Keras implementation, and deep reinforcement learning | Ch. 6, Secs. 2-3 and 5.1; Ch. 18, Secs. 1-10; coding context: Ch. 22, Secs. 2-3 | Coming soon |
Class examples
Datasets, code, and notebooks used for demonstrations and examples are collected here independently of the lecture schedule.
Variational inference · Variational autoencoders
Explore posterior approximation, the variational lower bound, and reparameterization. Train a small VAE and compare reconstruction with generation. Run variational_autoencoder_demo in MATLAB R2016b or later; no additional toolboxes are required.
Direct preference optimization · Preference learning
Compute the DPO loss for a preferred and a less-preferred response. Change the current-model probabilities, reference-model probabilities, and beta in the script to compare scenarios. Run DPO in MATLAB.
Final presentations
Register your group of up to four students now. Choose a presentation paper later, or propose another paper for approval.
Each group has 25 minutes total: approximately 15–20 minutes of prepared material and 5–10 minutes for discussion and questions. The split is flexible; questions are welcome throughout.
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 2 | Cautious Weight Decay |
| 1:25–1:50 p.m. | Group 15 | TimeBridge: Self-Supervised Video Representation Learning via Start-End Joint Embedding and In-Between Frame Prediction |
| 1:50–2:15 p.m. | Group 14 | Understanding the Learning Phases in Self-Supervised Learning via Critical Periods |
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 16 | Spectral Collapse Drives Loss of Plasticity in Deep Continual Learning |
| 1:25–1:50 p.m. | Group 11 | Platonic Transformers: A Solid Choice For Equivariance |
| 1:50–2:15 p.m. | Group 12 | Backward Oversmoothing: why is it hard to train deep Graph Neural Networks? |
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 9 | From Observations to States: Latent Time Series Forecasting |
| 1:25–1:50 p.m. | Group 18 | Retrieval Augmented Time Series Forecasting |
| 1:50–2:15 p.m. | Group 7 | Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? |
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 8 | How much can language models memorize? |
| 1:25–1:50 p.m. | Group 4 | Compute-Optimal Quantization-Aware Training |
| 1:50–2:15 p.m. | Group 5 | Log-Linear Attention |
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 6 | Let's (not) just put things in Context: Test-time Training for Long-context LLMs |
| 1:25–1:50 p.m. | Group 1 | LLMs Get Lost In Multi-Turn Conversation |
| 1:50–2:15 p.m. | Group 3 | Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains? |
| Time | Group | Paper |
|---|---|---|
| 1:00–1:25 p.m. | Group 17 | The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models |
| 1:25–1:50 p.m. | Group 10 | Flow Matching Policy Gradients |
| 1:50–2:15 p.m. | Group 13 | Generative Modeling via Drifting |
A pending paper does not change a group’s scheduled presentation time.
The presentation is worth 10% of the final course grade. The weights below apply to the presentation mark.
| Criterion | What we assess | Weight |
|---|---|---|
| Motivation & problem | Explain why the work matters and clearly define the problem it addresses. | 20% |
| Main idea & technical understanding | Explain how the approach works, with enough technical depth to demonstrate understanding. Establish the main ideas before finer details. | 35% |
| Findings & interpretation | Explain the main results, what we learn from them, and relevant limitations. | 20% |
| Clarity & organization | Teach the essential ideas clearly and selectively. Plan approximately 15–20 minutes of prepared material within the 25-minute session. | 15% |
| Discussion & questions | Allow approximately 5–10 minutes for discussion. Invite questions throughout and respond thoughtfully, demonstrating understanding. | 10% |
You are assessed on understanding and explanation, not on covering every detail. Discussion is assessed by how you facilitate and respond to questions, not by how many questions classmates ask.
Final project
Register your group of up to four students now. Choose a project later, or propose your own for approval. Find project descriptions, starter materials, report templates and the marking scheme on the project page.
Coursework
Individual assignment · 100 marks
Download the assignment PDF and both supplied Python files linked above. Keep data.py and model.py with your Question 7 code and use relative paths. Read the assignment instructions carefully; the Deliverables box at the end of each question specifies exactly what to submit.
WatIAM_A1.zip to the Assignment 1 Dropbox, replacing WatIAM with your Waterloo username. Combine the required files for Questions 1 and 7 in this single ZIP: Q1.py (or Q1.R/Q1.m), Q7.py, data.py, model.py, any helpers, Q7_results.csv, Q7_history.csv, Q7_test.csv, and one README covering commands, dependencies, and seeds. Do not submit separate Q1 and Q7 ZIPs or individual code files.For Question 1, use basic array operations and do not use automatic differentiation. For Question 7, use Python/NumPy and the supplied engine; write your own experiment loop, comparisons, logging, selection, and diagnostics, and do not call ready-made training, search, or early-stopping routines.
Generative AI use is permitted. After completing the assignment, open Learn and select Submit → Surveys → Assignment 1: Generative AI Use Survey. Everyone must complete the survey, including students who used no generative AI; in that case, answer No to all five questions. The survey is not anonymous so completion can be checked. Your survey responses do not affect your assignment grade. Open Learn.
You may be asked to complete a brief individual quiz or participate in an oral discussion about your assignment. This may include explaining your reasoning and solutions, answering related questions, or explaining and making small modifications to your submitted code. You must demonstrate your own understanding of all submitted work, including work developed with permitted generative AI assistance. These checks must be completed without generative AI or assistance from others. If you cannot demonstrate adequate understanding of your submitted work, your overall assignment mark may be reduced.
Submissions received up to 24 hours after the deadline receive a 10% penalty. Submissions more than 24 hours late receive a grade of zero.
Individual assignment · 100 marks
Download the assignment PDF and student starter ZIP. Extract the ZIP and keep the Q6 files together. The supplied data.py generates the data locally; no separate dataset download is required. The Deliverables box at the end of each question specifies what to submit.
WatIAM_A2.zip to the Assignment 2 Dropbox, replacing WatIAM with your Waterloo username. Include both Q3/ and Q6/, code and helpers, generated CSV tables and figures, and your own README.md with run commands, package versions and seeds. Use relative paths. Exclude environments, installed packages and trained-model checkpoints. Do not submit separate ZIPs for Q3 and Q6.Use Python with NumPy and Matplotlib for Q3; Q6 also uses PyTorch. Complete the five TODOs in Q6/model.py and write Q6/check_loss.py. The supplied Q6/run_q6.py handles data generation, training, evaluation and plotting. Equivalent independent code is permitted if it follows the PDF's protocol.
Submissions received up to 24 hours after the deadline receive a 10% penalty on the earned mark. Submissions more than 24 hours late receive a grade of zero.
Course tools
Use Piazza for announcements, lecture questions, clarifications, and class discussion.
Go to PiazzaUse Learn for graded work, submissions, private course documents, deadlines, and grades.
Open Waterloo LearnSign-in may be required. Students should use the course-specific links available in their enrolled Piazza and Learn accounts.
How the course runs
Public slides are posted with the tentative topics. Additional or restricted materials are shared through Learn.
Public assignment instructions are posted on this page. Submission links, feedback, and official grades are provided through Learn.
Use Piazza for course-content questions. Use a private Learn message or email for personal matters.
Consult Learn and Piazza for authoritative dates, room information, schedule changes, and deadlines.