Free Machine Learning Workshop: From Traditional ML to Graph Neural Networks

A math Discord server is organizing a completely free machine learning workshop series, covering the full knowledge path from traditional machine learning to deep learning and graph neural networks. For those who want to systematically get started with ML but struggle to find a suitable study group, this kind of environment—with instructors leading sessions, homework being reviewed, and community discussion—may be far more efficient than grinding through online courses alone.

Workshop Basics

According to a recruitment post published on Reddit, the event will take place on a Discord server called maths (discord.gg/maths), with lectures and/or lab sessions scheduled weekly. The server itself is a math community with a focused atmosphere, and participants are generally eager to help each other—a significant bonus given the heavy mathematical derivations involved in learning ML.

Participants need to prepare in advance: the course involves a considerable amount of coding exercises, and instructors will try to review each student's submitted code individually. This means it's not the kind of lecture where you just listen without practicing—it's a bootcamp-style program that requires actually writing code and submitting assignments.

Course Outline: Three Parts Covering the Full ML Stack

According to the schedule published in the post, the entire workshop is divided into three parts.

Part 1: Traditional Machine Learning. Starting from overall concepts and terminology, it covers linear models and gradient descent, supervised learning methods, unsupervised learning methods, and finally lands on machine learning best practices. This part addresses the foundational questions of "what is ML and how do you use it."

Part 2: Introduction to Deep Learning. This includes multilayer perceptrons and backpropagation, autoencoders, recurrent neural networks for sequence data, convolutional neural networks, attention mechanisms and Transformers, and signal processing. As you can see, this outline places Transformers within the framework of an introduction to deep learning rather than treating them as an isolated advanced topic—a fairly clear approach.

Part 3: Graph Machine Learning. The content covers traditional methods and feature engineering on graphs, shallow node embeddings, knowledge graph embeddings, graph neural networks, and subgraph sampling. The post also notes that more content will be added as the course progresses. Graph neural networks are increasingly widely applied in areas like recommendation systems and molecular modeling, and dedicating a separate section to them shows the organizers have a certain ambition for course depth.

Prerequisites: Python Is a Hard Requirement

The post lays out clear requirements for applicants, which boil down to four core items.

The most critical is Python programming ability, especially with common scientific computing libraries. You'll need to know numpy (official guides, books, and a roughly 1-hour video tutorial are all recommended) and matplotlib (official tutorial and w3schools tutorial provided), with Seaborn being optional but "much prettier." Additionally, some familiarity with Python object-oriented programming (at least knowing how to work with classes) will make things go more smoothly.

On the math side, Calculus I is a requirement, Calculus II is optional but helpful for understanding sequences; linear algebra only requires a basic level. The organizers admit they're still discussing exactly how deep the linear algebra needs to go, but they've made it clear they don't want to dig too deep in order to keep the entry barrier acceptable—a reasonable trade-off for a workshop aimed at a general audience.

Low Barrier, But High Commitment

The positioning of this workshop is worth mentioning. It's free, open, and anchored in a math community, but judging from the course outline, it goes all the way to Transformers and graph neural networks—it's not a superficial overview. The organizers are pushing forward with a "weekly lecture + lab" pace, and they'll also review code, which in practice demands sustained commitment from both instructors and students.

The original post specifically mentions that if you have questions, it's best to join the server and communicate directly with instructors rather than only asking in the post's comment section. This shows that the entire project depends on real-time interaction within the community—if you plan to participate, being prepared to invest fixed time each week and engage in discussions will yield far more than simply auditing.

Background Notes

Transformers and Attention Mechanisms: The core idea of attention mechanisms is to let a model, when processing information at a certain position, "attend" to content at other positions in the sequence and assign different weights based on relevance. Its basic form is scaled dot-product attention: map the input into three sets of vectors—Query, Key, and Value—measure similarity using the dot product of Q and K, normalize via softmax, and then compute a weighted sum of V. Transformers, meanwhile, completely abandon recurrent structures and rely solely on this attention mechanism plus feedforward networks, residual connections, and layer normalization to model sequences, thereby enabling highly parallel computation and becoming the foundational architecture for modern large models like BERT and GPT. Understanding this part is a prerequisite for comprehending today's mainstream NLP and multimodal models.

Graph Neural Networks (GNNs): Traditional neural networks assume data exists on regular grids or as independent samples, whereas graph data consists of nodes and edges, where each node has a variable number of neighbors with no meaningful ordering. GNNs solve this problem through a "message passing" mechanism: in each layer, a node first aggregates feature information from its neighbors, then combines it with its own representation and updates through a nonlinear transformation. After stacking multiple layers, each node can capture structural information from its multi-hop neighborhood. Common variants include GCN (simplified based on spectral approximation), GraphSAGE (sampling and aggregation, suitable for large graphs), and GAT (introducing attention-weighted neighbors). GNNs are widely used in recommendation systems, molecular property prediction, knowledge graph reasoning, and other scenarios, making them the core method in current graph machine learning.