These Chinese notes on Andrew Ng's Deep Learning Specialization were written from video subtitles

The notes follow the videos and subtitles, and the author says so

The README is in Chinese, titled as Chinese notes for the Deeplearning.ai deep learning course. It links the course page on NetEase's MOOC platform and an online reading site at ai-start.com/dl2017. The author, Dr. Huang Haiguang, states plainly that the notes are written from the video and subtitles, that they have no technical content of their own, and that they only require focus and rigor.

That is the right expectation to carry in. These are transcription-style study notes with formulas and the lecture structure, not a rewritten textbook. The author also says his level is limited and asks readers to email him about errors in formulas or algorithms, so a formula that looks wrong deserves a check against the video before you trust the note.

The author previously wrote personal notes for Andrew Ng's machine learning course, and the README links that other repository, fengdu78/Coursera-ML-AndrewNg-Notes. The notes are described as produced by a team of volunteers who help without pay and keep updating them. The README dates itself 2018-04-14, so the core text comes from that period.

The specialization has five courses and about three to four months of study

The README describes the target reader as a computer professional with basic programming knowledge, familiarity with Python and a basic grasp of machine learning, who wants to enter artificial intelligence. It describes the Deep Learning Specialization as covering convolutional neural networks (CNN), recurrent neural networks (RNN) and long short-term memory (LSTM), among other common network structures, tools and knowledge.

The course language is Python and the framework is Google's open-source TensorFlow. Andrew Ng is the instructor, and the two teaching assistants are from the Stanford computer science department. Finishing takes roughly three to four months depending on pace, and Coursera awards a Deep Learning Specialization completion certificate. The projects mentioned include healthcare, autonomous driving, natural language processing and music generation.

Because the notes are tied to a TensorFlow course from that era, framework code in the notes and the assignment folder reflects the course, not current library practice. For the underlying math and training ideas that matters little, and for API calls it matters a lot.

Course 1 builds logistic regression and vectorization before any deep network

The table of contents is the most useful part of the README. Course 1, Neural Networks and Deep Learning, runs four weeks. Week 1 is the introduction, including a Geoffrey Hinton interview. Week 2 covers programming basics: binary classification, logistic regression and its cost function, gradient descent, derivatives, computation graphs, vectorization, broadcasting in Python, notes on Python and numpy vectors, a quick tour of Jupyter and iPython notebooks, and an explanation of the logistic regression cost function.

Week 3 treats shallow neural networks: representation, computing the output, vectorizing across examples, activation functions, why nonlinear activations are needed, their derivatives, gradient descent, an optional lesson on backpropagation intuition, and random initialization. Week 4 moves to deep L-layer networks, forward and backward propagation, getting matrix dimensions right, why deep representations help, building blocks, parameters versus hyperparameters, and the link to the brain.

The order shows that vectorization and broadcasting come in week 2, before layers. If you are skimming for the numpy shape rules, lessons 2.15 and 2.16 are where they sit, and lesson 4.4 on matrix dimensions is the one to reread when a shape error appears.

Course 2 covers regularization, optimizers and batch normalization in three weeks

Course 2 is Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization. Week 1 covers practical aspects: train, dev and test sets, bias and variance, the basic recipe for machine learning, regularization and why it reduces overfitting, dropout and why it works, other regularization methods, normalizing inputs, vanishing and exploding gradients, weight initialization, numerical approximation of gradients, gradient checking and its implementation notes.

Week 2 is optimization algorithms: mini-batch gradient descent, exponentially weighted averages and their bias correction, momentum, RMSprop, Adam, learning rate decay and the local optima problem. Week 3 covers hyperparameter tuning, batch normalization and programming frameworks: the tuning process, picking an appropriate scale, the Pandas versus Caviar approach, normalizing activations, fitting Batch Norm into a network, why it works, Batch Norm at test time, softmax regression, training a softmax classifier, deep learning frameworks and TensorFlow.

A reader who already knows the basics can start here, since these are the topics that change training results in practice. Lessons 1.13 and 1.14 on gradient checking are the debugging tool for hand-written backpropagation.

Course 3 is about error analysis and data mismatch, not new models

Course 3, Structuring Machine Learning Projects, is the strategy course. Week 1 covers why ML strategy matters, orthogonalization, a single number evaluation metric, satisficing and optimizing metrics, train, dev and test distributions and their sizes, when to change the dev set, test set and metric, why human-level performance is used, avoidable bias, understanding and surpassing human-level performance, and improving model performance.

Week 2 covers error analysis, cleaning up incorrectly labeled data, building a first system quickly and then iterating, training and testing on different distributions, bias and variance with mismatched distributions, addressing data mismatch, and transfer learning.

This course contains no code-heavy derivations. It is useful as a decision guide for what to try next when a model underperforms, and the notes preserve that structure as a numbered list you can scan. The README's credits show the notes also cover courses 4 and 5, naming who drafted all of course 4 and which weeks of course 5.

The repository ships a PDF, Markdown, HTML, papers, comics and assignment code

The top-level entries are Deeplearning深度学习笔记v5.72.pdf, README.md, and folders for html, images, markdown, reference papers (参考论文), comics (漫画) and course assignment code (课程作业代码). The primary language is HTML, because of the rendered pages. The PDF at version 5.72 is the single file to download for offline reading.

The repository has no GitHub releases and no license file. The README states that the document is free, must not be used commercially, may be freely spread, and that reposts should credit github.com/fengdu78/deeplearning_ai_books. Treat that as the only usage terms. Copying the notes into a paid course or product is excluded by that wording.

The last push was on 5 October 2026, so the repository still receives changes, though the README itself is dated 2018. A video version of the course is linked on Bilibili, and a mobile-friendly notes site is at ai-start.com. The author lists a QQ group for machine learning, a WeChat official account and a Zhihu profile for contact.

Editorial conclusion

Read these notes if you can read Chinese and want a lecture-by-lecture companion to the Deep Learning Specialization, with the PDF for offline use and the folder of assignment code beside it. Do not expect English text, original technical content or a clear open license: the author says the notes follow the videos and subtitles, the README forbids commercial use, and the repository metadata lists no license.

Frequently asked questions

Official sources

1. fengdu78/deeplearning_ai_books on GitHub
2. Issues
3. README