CSE 455/555 Introduction to Pattern Recognition
In this updated Pattern Recognition course, we’ll learn not only by lectures and problem sets but by diving directly into the original research papers that shaped the field. Each week, you’ll read several papers—ranging from classic template-matching methods and statistical classifiers to modern deep-learning approaches—then discuss and implement key ideas.
Our in-class sessions will be a mix of presentations, guided code walkthroughs, and Q&A. You’ll write brief critical analyses for each paper, code up core algorithms in Python (using NumPy, scikit-learn or PyTorch), and compare performance on benchmark datasets. By the end of the semester, you’ll not only master pattern-recognition techniques but also develop the critical reading and implementation skills needed to engage with cutting-edge research.
Prerequisite: Pre-Requisites: CSE 250 or EAS 230 or EAS 240 or CSE 115 or EAS 999TRCP and EAS 305 or STA 301 and STA 301 or MTH 411; Computer Science, Computer Engineering, or Bioinformatics majors only. Students must complete a mandatory advisement session with their faculty advisor.
Resources: A curated PDF reading list will be provided; expect weekly code exercises and written critiques.
Instructor Information
Course Instructor: Jue Guo
- Research Area: Optimization for machine learning, Adversarial Learning, Continual Learning and Graph Learning
- Interested in participating in our research? Reach to me by email.
Course Outline and Logistics
Check out the course material under lecture notes.
Course Hours: MoWe 7:00 PM – 8:20 PM
Format: Remote
Term Dates: Jun 22, 2026 – Jul 31, 2026
Course Format
This is a paper-reading seminar. We read twelve landmark papers in sequence, tracing how modern deep learning was built — from vision foundations (AlexNet, ResNet), through the sequence-to-Transformer arc (Word2Vec, attention, the Transformer), into the large-language-model era (scaling laws, RLHF, chain-of-thought, and the latest reasoning models). You see how deep nets work on vision before diving into LLMs, and the back half builds the full modern LLM story end to end. Each session below includes a short reading guide telling you what to focus on and what question to keep in mind.
- Pre-class quiz. Each session is gated by a short online quiz (~10 minutes, mix of multiple-choice and short-answer) released 48 hours before class and due 30 minutes before it starts. Questions check basic comprehension: notation, headline claims, the experimental setup, and one "did you actually read it?" detail from the paper.
- In class. The instructor opens with a 15-min framing, then leads discussion focused on the items the quiz revealed students struggled with, followed by a deep dive into the hardest concept or derivation.
- No exams, no formal write-ups. Comprehension is verified continuously through the quizzes; the rest of your time goes into the final project.
Schedule
Twelve sessions over six weeks: three vision foundations → three sequence/Transformer → six on the LLM era. Each row gives the paper(s) and a short reading guide — what to focus on and a question to hold while you read.
| Arc | # | Date | Paper & reading guide |
|---|---|---|---|
| Vision foundations |
1 | Mon Jun 22 |
Kickoff + How to Read a Paper — Keshav 2007, How to Read a Paper Reading guide. Course logistics, expectations, and the three-pass reading method we use all term. Ask yourself: for a paper of your choice, can you do a 5-minute first pass and state its category, its contributions, and whether the claims look sound? |
| 2 | Wed Jun 24 |
Krizhevsky et al. 2012, ImageNet Classification with Deep CNNs (AlexNet) Reading guide. The result that launched the deep-learning era. Focus: the ingredients — ReLU, dropout, data augmentation, two-GPU training — and the size of the ImageNet error drop. Ask yourself: which ingredient mattered most, and which are still standard today? |
|
| 3 | Mon Jun 29 |
He et al. 2015, Deep Residual Learning (ResNet) Reading guide. How to train networks hundreds of layers deep. Focus: the degradation problem and why the skip connection $$F(x)+x$$ fixes it. Ask yourself: why is learning a residual easier than learning the full mapping? (Tie it back to vanishing gradients.) |
|
| NLP → Transformer |
4 | Wed Jul 1 |
Mikolov et al. 2013, Efficient Estimation of Word Representations (Word2Vec) Reading guide. Words become vectors. Focus: skip-gram vs. CBOW and the famous "king − man + woman ≈ queen" analogy. Ask yourself: what does direction and distance in embedding space encode, and how is it learned with no labels? |
| 5 | Mon Jul 6 |
Paired:
Sutskever et al. 2014, Sequence to Sequence Learning +
Bahdanau et al. 2015, NMT by Jointly Learning to Align and Translate (attention) Reading guide. From a single fixed context vector to attention. Focus: why one bottleneck vector fails on long sentences, and how attention lets the decoder look back at every encoder state. Ask yourself: what does a single attention weight actually represent? |
|
| 6 | Wed Jul 8 |
Vaswani et al. 2017, Attention Is All You Need Reading guide. The Transformer — recurrence removed entirely. Focus: scaled dot-product attention, multi-head attention, and positional encodings. Ask yourself: how does self-attention recover word order without an RNN? This architecture underlies every model after it. |
|
| LLM era | 7 | Mon Jul 13 |
Devlin et al. 2018, BERT Reading guide. Pretrain-then-fine-tune, the encoder side. Focus: masked language modeling and bidirectional context. Ask yourself: why does masking enable bidirectionality that a left-to-right language model cannot have? |
| 8 | Wed Jul 15 |
Brown et al. 2020, Language Models are Few-Shot Learners (GPT-3) Reading guide. Scale + in-context learning, the decoder side. Focus: few-/zero-shot prompting with no weight updates, at 175B parameters. Ask yourself: how is in-context learning different from fine-tuning, and why might scale alone produce it? |
|
| 9 | Mon Jul 20 |
Paired:
Kaplan et al. 2020, Scaling Laws for Neural LMs +
Hoffmann et al. 2022, Training Compute-Optimal LLMs (Chinchilla) Reading guide. How big a model, how much data? Focus: the power-law of loss vs. compute/params/data, and Chinchilla's correction that most LLMs were badly under-trained on tokens. Ask yourself: for a fixed compute budget, how should you split it between parameters and training tokens? |
|
| 10 | Wed Jul 22 |
Ouyang et al. 2022, Training LMs to Follow Instructions with Human Feedback (InstructGPT / RLHF) Reading guide. Turning a raw LM into something that follows instructions. Focus: the three stages — supervised fine-tuning → reward model → PPO. Ask yourself: why does a 1.3B InstructGPT beat 175B GPT-3 on human preference, and what is RLHF actually optimizing? |
|
| 11 | Mon Jul 27 |
Paired:
Wei et al. 2022, Chain-of-Thought Prompting +
Wei et al. 2022, Emergent Abilities of LLMs Reading guide. Reasoning from prompting, and abilities that appear with scale. Focus: how "let's think step by step" unlocks multi-step problems, and the claim that some skills emerge abruptly. Ask yourself: are emergent abilities a real phase change, or an artifact of how we measure them? (Hold both sides.) |
|
| 12 | Wed Jul 29 |
Reasoning models —
OpenAI 2024, o1 / Learning to Reason +
DeepSeek-AI 2025, DeepSeek-R1 Reading guide. Test-time compute and RL-trained reasoning. Focus: training models to generate long chains of thought via reinforcement learning, and "thinking longer" at inference. Ask yourself: how does R1's RL recipe differ from RLHF, and what does extra test-time compute buy you? Final-project lightning talks also run this final week — 5 min per student + Q&A. |
Grading
| Component | Weight & Details |
|---|---|
| Pre-class quizzes | 50% — one quiz per paper session (11 total, sessions 2–12), drop the lowest two |
| In-class participation | 15% — discussion quality & engagement |
| Final project | 35% — a build-it-from-scratch capstone, graded across six milestones (proposal → baseline checkpoint → draft → peer review → report + code → talk). See the Final Project page. |
Final Project
One capstone, done by everyone — no tracks. You will build a small language model from scratch, post-train it into an instruction-follower, and then investigate a sharp question about reasoning and scale, threading the whole back half of the syllabus (Transformer → GPT → scaling laws → RLHF → chain-of-thought). It is graded as a mini research paper across six milestones, with required baselines, ablations, multi-seed results, and a reproducible repo.
→ Read the full Final Project brief (stages, requirements, milestone deadlines, rubric, and starting points).
Note on Logistics
- A week-ahead notice for mid-term, based on the pace of the course.
- The logistic is subject to change based on the overall pace and the performance of the class.
Grading
The following is the outline of the grading:
Grading Rubric
This course is absolute grading, meaning no curve, as there is a certain standard we need to uphold for students to have a good knowledge of algorithm.
| Percentage | Letter Grade | Percentage | Letter Grade |
|---|---|---|---|
| 95-100 | A | 70-74 | C+ |
| 90-94 | A- | 65-69 | C |
| 85-89 | B+ | 60-64 | C- |
| 80-84 | B | 55-59 | D |
| 75-79 | B- | 0-54 | F |
AI Research Reading List
Neural Network Foundations
- Back-Propagation (Rumelhart, Hinton & Williams, 1986) – PDF · notes
- Dropout (Srivastava et al., 2014) – PDF · notes
- ReLU (Nair & Hinton, 2010) – PDF · notes
- Adam (Kingma & Ba, 2014) – arXiv:1412.6980 · notes
- Batch Normalization (Ioffe & Szegedy, 2015) – arXiv:1502.03167 · notes
- RNN Encoder–Decoder (Cho et al., 2014) – PDF · notes
Computer Vision
- AlexNet (Krizhevsky, Sutskever & Hinton, 2012) – NeurIPS 2012 PDF
- VGG (Simonyan & Zisserman, 2014) – arXiv:1409.1556
- ResNet (He et al., 2015) – arXiv:1512.03385
- Inception Net (Szegedy et al., 2015) – arXiv:1409.4842
- U-Net (Ronneberger et al., 2015) – arXiv:1505.04597
- Faster R-CNN (Ren et al., 2015) – arXiv:1506.01497
- YOLO (Redmon et al., 2016) – arXiv:1506.02640
- Mask R-CNN (He et al., 2017) – arXiv:1703.06870
- EfficientNet (Tan & Le, 2019) – arXiv:1905.11946
Natural Language Processing
- BERT (Devlin et al., 2018) – arXiv:1810.04805
- GPT-2 (Radford et al., 2019) – PDF
- RoBERTa (Liu et al., 2019) – arXiv:1907.11692
- T5 (Raffel et al., 2020) – arXiv:1910.10683
- GPT-3 (Brown et al., 2020) – arXiv:2005.14165
Graph Neural Networks
- Graph Attention Networks (Veličković et al., 2018) – arXiv:1710.10903
- Deep Graph Infomax (Veličković et al., 2019) – arXiv:1809.10341
Adversarial Machine Learning
- Adversarial Examples in the Physical World (Kurakin, Goodfellow & Bengio, 2016) – arXiv:1607.02533
- Fast Is Better Than Free (Wong, Rice & Kolter, 2020) – arXiv:2001.03994
Continual Learning
- Overcoming Catastrophic Forgetting (Kirkpatrick et al., 2017) – arXiv:1612.00796
- Learning without Forgetting (Li & Hoiem, 2018) – arXiv:1606.09282