Courses, papers, and projects by direction
Once a question interests you, use this catalog for courses, representative papers, code, and fuller paper lists. For a first introduction, read the directions overview, then return to choose material.
Computer vision: perception, representations, and image models
Kaggle Digit Recognizer: MNIST — Go from displaying data through training and validation to generating predictions and submitting results. Find beginner notebooks in the Code section and a concrete starting route in the guide. Code notebooks Hands-on route
Ultralytics YOLO Quickstart — Detect objects in your own photos with a pretrained model, then try small-dataset training and validation. A practical entry for a visual demo you can show.
Stanford CS231n — English lectures and slides on vision tasks, losses, training, and convolutional/visual architectures. Start there, then choose a topic from the schedule.
CNN Explainer — Polo Club — Interactive demonstration of convolutional networks, input images, and intermediate representations.
Computing Receptive Fields of Convolutional Neural Networks — Distill — An illustrated technical article explaining how convolutional architectures determine the input area visible at one position.
NLP and LLMs: concepts, use, implementation, and research
Hugging Face LLM Course — Lessons and code moving from language-model concepts to tokenizers, models, and data. Begin with a small working example.
Build a Large Language Model From Scratch — Sebastian Raschka — Open companion code and chapter material for an English book. Start with tokens, embeddings, attention, and training; the book is purchased separately.
Stanford CS224N — An English NLP course. Choose representations, attention, and language tasks to understand the questions and methods behind language models.
Stanford CS336: Language Modeling from Scratch, 2025 — English lectures, assignments, and videos connecting data, training, systems, and evaluation. Intended for people with DL and programming foundations. Official first lecture
Deep Dive into LLMs like ChatGPT — Karpathy — A broad-audience English video introducing LLMs, useful for an overall picture before implementation.
Multimodal models: connecting inputs and outputs
Vision Language Models Explained — Hugging Face — An English explanation of VLM components, use, and evaluation, useful when moving from LLMs into multimodal learning.
CLIP official project — Official code and paper. Begin with image–text correspondence, inputs and outputs, and zero-shot use before training details.
LLaVA: Visual Instruction Tuning — Author project page, diagrams, and paper. Examine model connections and instruction data to see how images enter language interaction.
Awesome Multimodal Large Language Models — A paper and resource list. Follow particular tasks to surveys and related work.
Generation, diffusion, and flow matching
Hugging Face Diffusion Course — Practical tutorials and notebooks. The first unit uses a small diffusion example for image generation; follow noising and denoising to learn the mathematics.
MIT: Introduction to Flow Matching and Diffusion Models, 2025 — English notes, lectures, and exercises connecting generation, flows, and diffusion mathematically. Requires probability and calculus.
Generative Modeling by Estimating Gradients of the Data Distribution — Yang Song — A technical blog explaining generative modeling through scores. Start with intuition and figures before derivations.
What are Diffusion Models? — Lilian Weng — A technical survey blog from 2021. Useful after the basics for connecting concepts and derivations.
World models: prediction, interaction, imagination, and decisions
World Models — David Ha and Jürgen Schmidhuber — An author article with visualizations and experiments, connecting compressed observations, predicted dynamics, and control in one system.
DreamerV3 — Danijar Hafner and colleagues — Paper, code, and demonstrations. Follow World Models into how a learned model serves behavior learning and how training and evaluation are organized.
Awesome World Models — JiahuaDong — A topic-organized paper list for investigating video, robotics, and other uses of world models.
RL: interaction, introductory texts, code, and depth
Hands-on Reinforcement Learning — A Chinese online book with code and videos. Begin with definitions and a small environment, then move from values and policies into deep RL.
Easy RL: The Mushroom Book — A Chinese community tutorial offering another explanation of unfamiliar RL concepts.
Gymnasium: Basic Usage — Official code tutorial showing the flow of observations, actions, rewards, and termination signals in small environments.
CleanRL — RL implementations and explanations. Pick an algorithm and follow its relatively concentrated training code beside the concepts.
Berkeley Deep Reinforcement Learning — Lectures, videos, and assignments. After RL foundations, explore model learning, policy optimization, and other topics.
Embodied AI: orientation, routes, papers, and practice
Lumina / Tianxing Chen: Embodied-AI-Guide — A Chinese embodied-AI guide. Get a technical map, then find routes, papers, tutorials, and projects by direction.
MIT Robotic Manipulation — Russ Tedrake — English course and textbook on geometry, perception, planning, and control, for understanding the pieces of manipulation.
MIT Underactuated Robotics — A control and robotics course for deeper study of dynamics, control, and decision-making by topic.
LeRobot — Open implementations, data, and documentation for robot learning. Start with demonstrations and model execution; real-robot work needs the corresponding hardware.
RoboTwin 2.0 — A simulation-data and evaluation project connecting tasks, data, training, and evaluation. Check listed GPU and environment requirements before installation and training.
ALOHA / ACT: Learning Fine-Grained Bimanual Manipulation — Project videos and paper connecting action data, imitation learning, and bimanual tasks. Begin with demonstrations and the method diagram.
Diffusion Policy — Author page and demonstrations applying generative modeling to action learning. Compare image diffusion with action sequences.
Open X-Embodiment — A cross-robot data and model project. Inspect sources, differences between robots, and how generalization is evaluated.
Specialized embodied-AI paper indexes
Awesome Humanoid Robot Learning — Yanjie Ze — A paper index to follow locomotion, manipulation, and other questions in humanoid learning.
Awesome RL-VLA — A paper list at the intersection of RL and VLA manipulation learning.
Awesome Efficient VLA — An index connecting VLA inference speed, deployment constraints, and model design.
Data Pyramid for Embodied Manipulation — Survey resources for understanding embodied-manipulation data through its sources and organization.
Agents: concepts, tutorials, projects, and evaluation
LLM Powered Autonomous Agents — Lilian Weng — An English technical blog from 2023 introducing LLM agents through planning, memory, tool use, and related components.
Hugging Face Agents Course — Official lessons and practice. Understand tools and execution, then build a small agent. Model-API exercises require accounts and may incur usage costs.
Hello Agents — Datawhale — A Chinese theory-and-practice tutorial. Begin with basic definitions and small implementations, then choose frameworks for the project.
LLM-Agent-Paper-List — A paper list maintained by survey authors. Inspect the classification first, then follow one question.
AgentBench — An agent-evaluation project. Read task setups, success criteria, and failures to understand how completion is assessed.
SWE-bench — Software-task evaluation. Inspect the relationship between real issues, tests, and results; compare tools, budgets, and settings as well as scores.
Efficiency, inference, and hardware–software design
Machine Learning Systems — A systems textbook connecting models to data, software, hardware, and deployment. Consult the topic behind a current bottleneck.
MIT 6.5940: TinyML and Efficient Deep Learning Computing, 2024 — English lectures, videos, and material for systematic study of quantization, compression, and efficient computation.
Making Deep Learning go Brrrr From First Principles — Horace He — An English engineering blog explaining why high nominal compute does not guarantee speed, through computation, memory, and execution overhead.
How To Scale Your Model — An open technical book on distributed training, inference, and hardware constraints. Choose chapters around a bottleneck.
PyTorch Performance Tuning Guide — Official performance guidance. Select a relevant change and measure before and after with the same workload.
AI4X: scientific data and tasks
DeepChem — Open tools and tutorials. Pick a drug or molecule task with clear inputs, labels, and evaluation to understand domain-specific data.
Therapeutics Data Commons — Data and evaluation for therapeutic and biological tasks. Read task definitions and splits before investigating models.
MONAI Tutorials — Medical-imaging tutorials. Start with processing and segmentation examples to connect images, annotations, and evaluation.
DeePMD-kit — A molecular-dynamics and ML project for materials and molecular simulation, requiring relevant physics and chemistry.
FinRL — A financial RL project. Inspect environment, splits, and backtesting, particularly temporal ordering, transaction costs, and evaluation conditions.
Awesome AI for Science — A cross-disciplinary collection of tools, papers, and projects. Find specific work by subject.
Automated research as a system to inspect
autoresearch — Karpathy — Automatic iteration on training experiments. Inspect how goals, editable scope, result records, and evaluation form an experiment loop.
The AI Scientist — Sakana AI — Paper, code, and documentation showing a system for ideas, experiments, and writing. Follow how each stage is evaluated.
Graph learning and time-series forecasting
Stanford CS224W: Machine Learning with Graphs — An English graph-learning course. Learn representations and typical tasks, then investigate your own relational-data question.
Forecasting: Principles and Practice, third edition — An open forecasting textbook using R. Learn trends, seasonality, prediction, and evaluation; Python users can start with concepts and figures.
tsai — PyTorch/fastai time-series tools and examples, useful for practical approaches once you have a task and data.