Pattern Recognition 2025

I teach a pattern recognition course for BME Ph.D. students. In this course, I will cover some advanced concepts of AI.

Description

This course aims to equip Ph.D. students with a deep understanding of modern pattern recognition theory and cutting-edge innovation capabilities. The curriculum covers five key areas: fundamental theories, discriminative models, reinforcement learning, generative models, and emerging topics such as robustness and fairness.

The course emphasizes the integration of theory and practice, requiring students to master end-to-end research methodologies—from mathematical derivation to engineering implementation—while fostering cross-domain innovation. Upon completion, Ph.D. students will possess the academic competence to publish in top-tier conferences.

This course takes place in JA302 every Thursday afternoon (14:15-17:15) during the first semester of the 2025-2026 academic year.

Expectations

Prerequisites

• Ph.D. student in a relevant field
• Python proficiency is required
• Strong foundation in mathematics and AI fundamentals

Course Requirements

• Paper presentation (in pairs of 2)
• Final deliverable: research draft paper or scientific blog post

Welcome & What’s Ahead

Here is some examples of what will be covered in this course.

Representation Learning

Instead of a human engineer manually designing features (like edges, shapes, or specific keywords), the machine learns to identify the most useful ways to represent the data for the problem at hand.

Here is an example of the two-moon representation:

This figure is interactive and was created using Plotly. It’s great, isn’t it?

Discriminative Models

It directly learns to map input features (\(X\)) to output labels or classes (\(Y\)) by modeling the conditional probability \(P(Y|X)\).

Here’s an example using Support Vector Regression on the famous Iris dataset:

Try to rotate this figure and observe how the surface is fitted to the dots.

Generative Models

It is a statistical model that learns the underlying probability distribution of the data with the goal of understanding how the data is “generated.”

Here’s an interactive example based on flow matching showing the transformation from a Gaussian distribution to a more complex multimodal distribution:

I like the idea of flow matching as it is particularly useful in generative modeling, where we want to learn to transform noise into realistic data samples while maintaining the ability to compute exact likelihoods and perform efficient sampling.

Schedule

Module 1: Basic Theory

Week	Date	Lesson 1	Lesson 2	Lesson 3	Materials
1	09/11/2025	Lecture	Lecture	Lecture	Slides
2	09/18/2025	Lecture	🧑🏼‍🏫👩🏼‍🏫Pres. I	Coding I	Slides, Code
3	09/25/2025	Reading1	Reading2	Reading3
4	10/02/2025	Reading4	Reading5	Reading6

¹ Understanding black-box predictions via influence functions, 2017 Read paper
² Matching networks for one shot learning, 2016 Read paper
³ Understanding deep learning requires rethinking generalization, 2016 Read paper
⁴ Deep double descent: Where bigger models and more data hur, 2021 Read paper
⁵ Auto-Encoding Variational Bayes, 2013 Read paper
⁶ Visualizing the loss landscape of neural nets, 2018 Read paper

Those classic papers are fundementals to this research fiel and a must know for a qualified Ph.D. student for AI related majors.

Module 2: Discriminative Models

Week	Date	Lesson 1	Lesson 2	Lesson 3	Materials
5	10/09/2025	Lecture	Lecture	Lecture	Slides
6	10/16/2025	Lecture	🧑🏼‍🏫👩🏼‍🏫Pres. II	Coding II	Slides, Code
7	10/23/2025	Reading7	Reading8	Reading9

⁷ A simple framework for contrastive learning of visual representations, 2020 Read paper
⁸ An image is worth 16x16 words: Transformers for image recognition at scale, 2020 Read paper
⁹ Flamingo: a visual language model for few-shot learning, 2022 Read paper

Module 3: Reinforcement Learning

Week	Date	Lesson 1	Lesson 2	Lesson 3	Materials
8	10/30/2025	Lecture	Lecture	Lecture	Slides
9	11/06/2025	Lecture	🧑🏼‍🏫👩🏼‍🏫Pres. III	Coding III	Slides, Code
10	11/13/2025	Reading10	Reading11	Reading12

¹⁰ Proximal policy optimization algorithms, 2017 Read paper
¹¹ BOHB: Robust and efficient hyperparameter optimization at scale, 2018 Read paper
¹² Deep reinforcement learning from human preferences, 2017 Read paper

Module 4: Generative Models

Week	Date	Lesson 1	Lesson 2	Lesson 3	Materials
11	11/20/2025	Lecture	Lecture	Lecture	Slides
12	11/27/2025	Reading13	Reading14	Reading15
13	12/04/2025	Lecture	🧑🏼‍🏫👩🏼‍🏫Pres. IV	Coding IV	Slides, Code

¹³ Pixel Recurrent Neural Networks, 2016 Read paper
¹⁴ Deep Image Prior, 2018 Read paper
¹⁵ Scaling Rectified Flow Transformers for High-Resolution Image Synthesis, 2024 Read paper

Module 5: Emerging Topics

Week	Date	Lesson 1	Lesson 2	Lesson 3	Materials
14	12/11/2025	Lecture	Lecture	Lecture	Slides
15	12/18/2025	Lecture	🧑🏼‍🏫👩🏼‍🏫Pres. V	Coding V	Slides, Code
16	12/25/2025	Reading16	Reading17	Reading18
17	01/01/2026			📆Final

¹⁶ Highly accurate protein structure prediction with AlphaFold, 2021 Read paper
¹⁷ Learnable latent embeddings for joint behavioural and neural analysis, 2023 Read paper
¹⁸ Generative models improve fairness of medical classifiers under distribution shifts, 2024 Read paper

Course Components

Presentation Sections

• Slots: 5 sessions (~10 slots, 20 students) (🧑🏼‍🏫👩🏼‍🏫Pres.) throughout the semester
• Group Size: Maximum 2 people per group
• Duration: 15 minutes presentation + 5 minutes Q&A
• Format: Interactive presentations with Q&A sessions

Coding Sections

• Slots: 5 sessions (I will prepare the materials) throughout the semester
• Hands-on Programming: Practical implementation of algorithms
• Languages: Python
• Environment: Jupyter notebooks, potentially with Google Colab

Note: These coding sessions are optional. Students are not required to stay in the classroom during these sessions.

Reading Sections

• Slots: 18 sessions throughout the semester I will provide a list of reading papers, and you should read them independently at your own pace.

Note: Please submit annotated PDFs of at least two papers to demonstrate your in-depth reading and analysis alongside your final project submission.

Final Project

• Due Date: Week 17, January 1, 2026, 23:59 China Standard Time (CST)
  
• Submission: Please submit the following via email to zejuli@fudan.edu.cn: – 1. Your final project draft – 2. At least two annotated reading papers

Note: The final project can be completed in one of the following formats:

• Scientific Blog: Present important course-related concepts in a rigorous, comprehensive, and pedagogically meaningful manner. Examples: Distill, ICLR Blog.

• Research Report: Conduct innovative research in the field of pattern recognition, with results that meet the standards of top-tier AI conference workshops.

Grading Policy

Component	Weight	Description
Presentations	30%	Clarity, relevance, and presentation layout
Final Project	60%	Research depth and technical writing
Participation	10%	Demonstrate engagement with paper reading

Recommended Reading Materials

Core Pattern Recognition & Machine Learning

• Pattern Recognition and Machine Learning - Christopher Bishop’s comprehensive textbook on pattern recognition
• 机器学习 - 周志华’s foundational machine learning textbook (in Chinese)

Deep Learning

• Deep Learning - Ian Goodfellow, Yoshua Bengio, and Aaron Courville’s authoritative deep learning textbook
• Dive into Deep Learning - Interactive deep learning book with practical implementations

Mathematical Foundations

• Mathematics for Machine Learning - Essential mathematical concepts for machine learning
• Machine Learning: A Probabilistic Perspective - Kevin Murphy’s detailed treatment of machine learning from a probabilistic viewpoint

Medical Imaging & Domain-Specific Applications

• Generative Machine Learning Models in Medical Image Computing - Specialized book on generative models in medical imaging
• Machine Learning in MRI - Focused on machine learning applications in magnetic resonance imaging