Rutgers University · Fall 2026

Deep Learning

01:198:462:02 · Foundations, models, and AI-assisted implementation

A rigorous, demo-first introduction to deep learning. We begin with working models, unpack the mathematics and design choices behind them, and build toward an individual final project.

Meeting
Fridays
12:10–3:10 p.m.
Location
LSH-A143
Livingston Campus
Instructor
Chengzhi Mao
cm1838@cs.rutgers.edu
Assessment
Project 50 points
Final exam 50 points

How the course works

Class time is for seeing models work, asking why they work, and practicing the reasoning needed to build and debug them.

See the demo

Each unit starts with a compact working example in PyTorch: data, model, loss, training loop, and observable behavior.

Unpack the ideas

We then trace tensor shapes, objectives, gradients, architectural choices, and failure modes using equations and diagrams.

Build and explain

Students extend the demo, use outside resources to fill gaps, and learn to explain what the code is doing—not merely run it.

AI coding is part of the course

Students are expected to learn an AI coding workflow using tools such as Claude Code or Codex. These tools may help scaffold, test, and debug project code, but students must verify the output and be able to explain every submitted component. No paid AI account is required.

Prerequisites

  • Calculus: derivatives, chain rule, and gradients, including matrix inputs
  • Linear algebra: vectors, matrices, norms, and the basic idea of SVD
  • Probability: conditional probability and Bayes' rule
  • Python: reading, writing, and debugging small programs

What to bring

  • A laptop for regular class demos and coding activities
  • A paper notebook for derivations, diagrams, and review
  • Curiosity and a willingness to inspect model failures

No previous deep learning course is assumed.

Learning outcomes

By the end of the course, students should be able to:

  1. Translate a learning problem into data, parameters, a model, a loss, and an evaluation plan.
  2. Trace forward and backward computation through linear models, MLPs, CNNs, RNNs, and transformers.
  3. Read and debug a compact PyTorch training pipeline, including tensor shapes and optimization behavior.
  4. Explain self-attention, causal language modeling, and the core structure of a GPT-style model.
  5. Compare VAE and GAN objectives, strengths, failure modes, and appropriate uses.
  6. Design, implement, evaluate, and clearly present an individual deep learning project.

Grading

The course grade is based on two assessments totaling 100 points.

There is no midterm exam.

50 points

Individual project

Build a deep learning system or experiment with a visible demo, evidence-based evaluation, and a concise explanation of what was learned.

Presentations: Weeks 11 and 12. Your final artifact is due before your assigned presentation.

50 points

Final exam

A cumulative, handwritten exam on the last class meeting: Friday, December 4, 2026, 12:10–3:10 p.m.

Closed book and no electronics. Expect conceptual questions, calculations, tensor shapes, diagrams, and pseudocode.

Project rubric. AI-generated rubric feedback may be used as an advisory check; the instructor assigns the final score.
Criterion Points Evidence expected
Technical depth 15 Appropriate deep learning or generative AI methods, implemented and explained correctly.
Evaluation and reproducibility 10 Clear setup, sensible metrics or comparisons, runnable code, and honest limitations.
Novelty and demonstration 10 An interesting question, extension, failure analysis, or visualization beyond copying a tutorial.
Presentation and understanding 15 A focused live explanation and convincing answers to questions about design and code.
Total 50 Individual work; material AI assistance must be disclosed.

Week-by-week plan

The sequence moves from machine learning basics through discriminative, sequential, attention-based, and generative models. Topics may be adjusted slightly to match class progress.

Week Date Focus Demo and in-class practice
01 Fri Sep 4 Machine learning basicsdatalossgeneralization Train a first classifier; identify inputs, parameters, objective, train/validation/test splits, and sources of error.
02 Fri Sep 11 Linear neural networkslinear regressionsoftmax Implement vectorized linear and softmax models; connect logits, cross-entropy, and one gradient step.
03 Fri Sep 18 Multilayer perceptronsactivationsbackprop Turn a linear model into an MLP; trace shapes and gradients; examine underfitting and overfitting.
04 Fri Sep 25 Convolutional neural networkskernelsreceptive fields Build an image classifier and visualize learned filters, feature maps, padding, stride, and pooling.
05 Fri Oct 2 Modern CNNsresidual blocksnormalization Compare plain and residual networks; use transfer learning and diagnose training behavior.
06 Fri Oct 9 Recurrent neural networkssequence modelingLSTM / GRU Train a small sequence model; unroll it through time and inspect vanishing or exploding gradients.
07 Fri Oct 16 Attention and transformersQ / K / Vself-attention Compute attention by hand and in code; assemble a transformer block with positional information.
08 Fri Oct 23 Language models and GPTtokenizationcausal masking Train a tiny next-token model; inspect tokens, causal attention, generation, and decoding choices.
09 Fri Oct 30 OptimizationSGDmomentumAdam Compare optimizers and learning-rate schedules; read loss curves and debug unstable training.
10 Fri Nov 6 VAE and GANlatent variablesadversarial learning Compare reconstructions and samples; trace reparameterization, the ELBO, GAN objectives, and training failure modes.
11 Fri Nov 13 Project presentations Iindividual project Live demos, technical explanations, questions, and rubric-based feedback.
12 Fri Nov 20 Project presentations IIindividual project Live demos, technical explanations, questions, and rubric-based feedback.
13 Wed Nov 25 Final exam review and go-throughFriday classes meet Wednesday Connect the full course sequence; work through representative calculations, diagrams, and pseudocode questions.
14 Fri Dec 4 Final exam50 pointshandwritten Cumulative in-class exam. Closed book; no phones, laptops, tablets, smartwatches, AI tools, or other electronics.

The Rutgers Fall 2026 calendar designates Wednesday, November 25 as a Friday-class day. The December 4 in-class final is the planned last meeting; because Rutgers ordinarily schedules final exams during the official exam period, this timing is subject to the required department/dean approval. If the plan is not approved, the official Rutgers exam schedule will control.

Essential policies

The shortest version: use tools honestly, own your work, and ask for help early.

AI use and integrity

AI coding tools are permitted and expected for learning and project development. Disclose material use, verify all output, and never submit work you cannot explain. AI tools are prohibited during the final exam. Rutgers academic integrity standards apply.

Attendance and preparation

The course is face-to-face and class meetings include demos and practice that are difficult to reproduce from slides alone. Attend prepared to code, take notes, ask questions, and participate.

Access and accommodations

Students seeking disability-related accommodations should contact the Rutgers Office of Disability Services and share the approved letter with the instructor as early as possible.