What do you think?
Rate this book
Machine Learning Yearning
AI, machine learning, and deep learning are transforming numerous industries. But building a machine learning system requires that you make practical decisions:
Should you collect more training data?
Should you use end-to-end deep learning?
How do you deal with your training set not matching your test set?
and many more.
Historically, the only way to learn how to make these "strategy" decisions has been a multi-year apprenticeship in a graduate program or company. This is a book to help you quickly gain this skill, so that you can become better at building AI systems.
Should you collect more training data?
Should you use end-to-end deep learning?
How do you deal with your training set not matching your test set?
and many more.
Historically, the only way to learn how to make these "strategy" decisions has been a multi-year apprenticeship in a graduate program or company. This is a book to help you quickly gain this skill, so that you can become better at building AI systems.
118 pages, ebook
131 people are currently reading
2377 people want to read
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 30 of 48 reviews
December 11, 2018
My favorite course from the Deep Learning Specialization on Coursera was "Structuring Machine Learning Knowledge" because it contained practical insights that were difficult to find elsewhere. Deep Learning Yearning contains much of the same information as that course, much of it expounded upon, and in a format that is easy to share with teammates and collaborators. As such, I think it is a valuable contribution to the field and deserves a spot on practitioner's bookshelves. Andrew Ng's writing style manages to be smooth and informal while maintaining precision of terminology. Another thought I had about the Coursera specialization was that Ng seemed to be using it as a platform to introduce standards and conventions to a rapidly evolving discipline that didn't really have well-established ones. His book appears to continue this trend - something I find most welcome. Highly recommend.
December 30, 2021
Andrew Ng is giving practical advice to the ML engineer through his experience at Google Brian, Baidu and teaching (Stanford and Coursera). It's a light technical book, giving succinct technical advice from someone being in the field...rules of thumbs, tricks, layman advice from lots of practice, trial and error.
Read
April 12, 2019تصمیم داریم کتاب یادگیری ماشین زیر را نوشته شده توسط یکی از بزرگان هوش مصنوعی جهان هستش به صورت گروهی به زبان فارسی به طور رایگان ترجمه و منتشر کنیم کسی از دوستان تمایل داشت در این کار همکاری کنه لطفا بهم در خصوصی پیام بده
January 2, 2021
classic ML reference book from Andrew Ng, geared towards ML practitioners
December 15, 2019
I've just read the "Takeaways" page in the end and it's enough for now.
The book deals with classical machine learning and not convolutional neural networks, so I'll get to read it fully someday later when it's out of beta and I have time to study it and Bishop's Machine learning book.
The book deals with classical machine learning and not convolutional neural networks, so I'll get to read it fully someday later when it's out of beta and I have time to study it and Bishop's Machine learning book.
April 16, 2018
Read the draft chapters and I can't wait for the final version to be released.
July 6, 2019
Lightweight, technical and practical. Highly recommended.
June 22, 2020
4 stars because it is a rehashed version of the ' Machine Learning Pipeline' Course on Coursera
January 24, 2020
The e-book is full of high-level descriptions, which should satisfy AI project managers who would like to have a quick introduction into the basics of ML. For the more technical readers, I would instead propose to have a look into Andrew's ML Coursera material.
Overall, the material introduces lots of practical approaches, mainly for diagnosing errors in the ML systems. Still, I think that some of the presented concepts, such as "eyeball dev set" and "unavoidable bias" are not that common in the industry and could have been omitted.
Overall, the material introduces lots of practical approaches, mainly for diagnosing errors in the ML systems. Still, I think that some of the presented concepts, such as "eyeball dev set" and "unavoidable bias" are not that common in the industry and could have been omitted.
December 17, 2022
This is a review of a draft of this book.
This book provides design patterns for error analysis of machine learning models.
I wouldn't rely on this book for machine learning theory of course, but most machine learning courses/books focus heavily on theory & standard algorithm implementations, neglecting the bread and butter of ML i. e. error analysis. So this book is a good supplement to the other resources.
Will be keeping this book close at hand for future reference!
(Less important, but this book could use a glow-up in terms of formatting and the cringey cover)
This book provides design patterns for error analysis of machine learning models.
I wouldn't rely on this book for machine learning theory of course, but most machine learning courses/books focus heavily on theory & standard algorithm implementations, neglecting the bread and butter of ML i. e. error analysis. So this book is a good supplement to the other resources.
Will be keeping this book close at hand for future reference!
(Less important, but this book could use a glow-up in terms of formatting and the cringey cover)
April 22, 2021
Interesting book for machine learning practitioners. Gives some general guidelines without being too technical. However if you are interested in learning about, or getting started on Machine Learning and/or Deep Learning, this is not the book for it.
January 9, 2020
More for those currently on the practical side creating something. Bits and pieces also useful for those looking to understand what AI is better.
August 24, 2023
If you are familiar with the world of Machine Learning, you better be aware of the great
Andrew Ng, who revolutionized the educational era with his exceptional teachings and researches in the field of AI, ML, NLP and more. I recently got a chance to read this book “Machine Learning Yearning” by him, which shows the technical strategy of AI Engineers in the era of deep learning.
It is divided into various segments, focusing on each aspects of model building and selection of right kind of data. Also, after each section, author has written a bonus chapter of “takeaways” as a summary of the topic, which will surely help budding team members to revise the important findings. I really appreciate the way Andrew has explained with very basic examples of simple model building, clarifies all the problems while choosing correct kind of data, how to bifurcate them, how to balance the evaluation metrics, and how to decide the optimized approach. Any beginner in Machine Learning can follow the details along very easily.
Following are the 15 major points which I find helpful:
- The book starts with the need of fashioning training and test sets of real world, and how it can be segregated with respect to abundant of distribution patterns. How even 0.01% of improvement in the performance can boost the moral of developer, it all depends on the size of data and fairly it has been distributed, but at the same time he stated that “there is no need to have excessively large dev/test sets beyond what is needed to evaluate the performance of your algorithms”.
- The importance of evaluation metric to be optimized is clearly shown through various examples.
There’s no shame in changing the dev/test set or even the metric, if the classifiers are leading the the project in wrong directions, and unnecessarily wasting the time of efficient team.
- Don’t start off trying to design and build the perfect system. Instead, build and train a basic system quickly and work on the iteration part to make the model better at each step. Two hours of manual work on error analysis can save you a month of wasted effort, as it provides a quantitative basis on which to decide whether to make the investment of choosing right thing.
- Error analysis does not produce a rigid mathematical formula that tells you what the highest priority task should be. You also have to take into account how much progress you expect to make on different categories and the amount of work needed to tackle each one.
- Consider splitting the dev set into an Eyeball dev set, which you will manually examine, and a Blackbox dev set, which you will not manually examine. If performance on the Eyeball dev set is much better than the Blackbox dev set, you have overfit the Eyeball dev set and should consider acquiring more data for it.
- Implement practical techniques to balance the trade-off between variance and bias.
- Try to resolve the noise in training curve by either train randomly chosen training set by sampling with replacement from original set, or choose a ‘balanced’ subset.
- Progress is usually slower on problems where machines already surpass human-level performance, while progress is faster when machines are still trying to catch up to humans.
- The clear picture of how to decide whether to use all or your data or include inconsistent data is stated by minor facts of various common models like cat detection, prediction of house prices etc.
- When synthesizing data, put some thought into whether you’re really synthesizing a representative set of examples. Try to avoid giving the synthesized data properties that makes it possible for a learning algorithm to distinguish synthesized from non-synthesized examples.
- Even though end-to-end learning has seen many successes, it is not always the best approach
- The examples of choosing pipeline components are more clearer by the statement — If you are able to take a complex task, and break it down into simpler sub-tasks, then by coding in the steps of the sub-tasks explicitly, you are giving the algorithm prior knowledge that can help it learn a task more efficiently.
- When you have the right (input, output) labeled pairs, you can sometimes learn end-to-end even when the output is a sentence, an image, audio, or other outputs that are richer than a single number.
- What if each individual component of your ML pipeline is performing at human-level performance or near-human-level performance, but the overall pipeline falls far short of human-level? This usually means that the pipeline is flawed and needs to be redesigned. Error analysis can also help you understand if you need to redesign your pipeline.
So, don’t just stick to these summary of the whole book but read the entire book to build a better team, and of course model.
Andrew Ng, who revolutionized the educational era with his exceptional teachings and researches in the field of AI, ML, NLP and more. I recently got a chance to read this book “Machine Learning Yearning” by him, which shows the technical strategy of AI Engineers in the era of deep learning.
It is divided into various segments, focusing on each aspects of model building and selection of right kind of data. Also, after each section, author has written a bonus chapter of “takeaways” as a summary of the topic, which will surely help budding team members to revise the important findings. I really appreciate the way Andrew has explained with very basic examples of simple model building, clarifies all the problems while choosing correct kind of data, how to bifurcate them, how to balance the evaluation metrics, and how to decide the optimized approach. Any beginner in Machine Learning can follow the details along very easily.
Following are the 15 major points which I find helpful:
- The book starts with the need of fashioning training and test sets of real world, and how it can be segregated with respect to abundant of distribution patterns. How even 0.01% of improvement in the performance can boost the moral of developer, it all depends on the size of data and fairly it has been distributed, but at the same time he stated that “there is no need to have excessively large dev/test sets beyond what is needed to evaluate the performance of your algorithms”.
- The importance of evaluation metric to be optimized is clearly shown through various examples.
There’s no shame in changing the dev/test set or even the metric, if the classifiers are leading the the project in wrong directions, and unnecessarily wasting the time of efficient team.
- Don’t start off trying to design and build the perfect system. Instead, build and train a basic system quickly and work on the iteration part to make the model better at each step. Two hours of manual work on error analysis can save you a month of wasted effort, as it provides a quantitative basis on which to decide whether to make the investment of choosing right thing.
- Error analysis does not produce a rigid mathematical formula that tells you what the highest priority task should be. You also have to take into account how much progress you expect to make on different categories and the amount of work needed to tackle each one.
- Consider splitting the dev set into an Eyeball dev set, which you will manually examine, and a Blackbox dev set, which you will not manually examine. If performance on the Eyeball dev set is much better than the Blackbox dev set, you have overfit the Eyeball dev set and should consider acquiring more data for it.
- Implement practical techniques to balance the trade-off between variance and bias.
- Try to resolve the noise in training curve by either train randomly chosen training set by sampling with replacement from original set, or choose a ‘balanced’ subset.
- Progress is usually slower on problems where machines already surpass human-level performance, while progress is faster when machines are still trying to catch up to humans.
- The clear picture of how to decide whether to use all or your data or include inconsistent data is stated by minor facts of various common models like cat detection, prediction of house prices etc.
- When synthesizing data, put some thought into whether you’re really synthesizing a representative set of examples. Try to avoid giving the synthesized data properties that makes it possible for a learning algorithm to distinguish synthesized from non-synthesized examples.
- Even though end-to-end learning has seen many successes, it is not always the best approach
- The examples of choosing pipeline components are more clearer by the statement — If you are able to take a complex task, and break it down into simpler sub-tasks, then by coding in the steps of the sub-tasks explicitly, you are giving the algorithm prior knowledge that can help it learn a task more efficiently.
- When you have the right (input, output) labeled pairs, you can sometimes learn end-to-end even when the output is a sentence, an image, audio, or other outputs that are richer than a single number.
- What if each individual component of your ML pipeline is performing at human-level performance or near-human-level performance, but the overall pipeline falls far short of human-level? This usually means that the pipeline is flawed and needs to be redesigned. Error analysis can also help you understand if you need to redesign your pipeline.
So, don’t just stick to these summary of the whole book but read the entire book to build a better team, and of course model.
October 24, 2024
After taking Andrew's Machine Learning Specialization class, this book is a perfect next step for me. It offers clear and practical advice for anyone navigating machine learning and AI projects, even if you are not a technical expert. It focuses on the importance of strategy—like how to set up the right development and test sets, and the value of a single evaluation metric. It’s not just about algorithms; it’s about how to think and approach real-world machine learning problems. One of the standout features of the book is its use of a simple, relatable example: a startup working on a platform for cat pictures. Through this, he breaks down key topics like error analysis, overfitting, and data issues in a way that’s easy to follow. Chapters like those on setting dev/test sizes and adapting metrics are particularly helpful if you're aiming to improve your project iteration process.
As Albert Einstein said, The definition of genius is taking the complex and making it simple. I really appreciated his style of teaching.
As Albert Einstein said, The definition of genius is taking the complex and making it simple. I really appreciated his style of teaching.
October 26, 2019
This can be a good book to always have on the desk, why? It lists down most of the quick diagnosing things in one place, so we can look at it in the moment of confusion. If you're an ML practitioner it doesn't tell anything that is new related to algorithms or anything like that but it helps one structure his/her ml project in a coherent manner. And small bite sized chapters, much like his lectures help us to search exactly for the problem we are facing and take an action that gives the best returns.
This book will help one brush his/her thoughts in planning properly for a project and give better project completion estimates, since we have to include for things like time for error analysis, time to reclassify misclassified data etc.
I wish the scope of the book was more with respect to deep learning, like for example in the chapter titled learning from curves a great deal of important things like loss vs epoch logs have been not discussed.
This book will help one brush his/her thoughts in planning properly for a project and give better project completion estimates, since we have to include for things like time for error analysis, time to reclassify misclassified data etc.
I wish the scope of the book was more with respect to deep learning, like for example in the chapter titled learning from curves a great deal of important things like loss vs epoch logs have been not discussed.
December 29, 2020
I fail to see who is the target reader here and what is the goal of this book.
• Are you new to ML? This book is not for you.
• Did you complete one of Andrew Ng’s online courses? The content is pretty much the same (even the examples are the same) so you won’t learn anything new.
• Are you experienced ML practitioner? It lists down quick diagnosis ways so I guess it could serve as a cheat sheet but that’s pretty much it.
I am confused whether this is an MVP of the future book but I can’t say I found much value in reading this.
• Are you new to ML? This book is not for you.
• Did you complete one of Andrew Ng’s online courses? The content is pretty much the same (even the examples are the same) so you won’t learn anything new.
• Are you experienced ML practitioner? It lists down quick diagnosis ways so I guess it could serve as a cheat sheet but that’s pretty much it.
I am confused whether this is an MVP of the future book but I can’t say I found much value in reading this.
January 25, 2019
Hallmark of Andrew's teachings is the ability to present the most complex concepts in the simplest possible terms without losing the essence of the lesson. This book is an exemplar of that ability. This book is like a timeless cookbook for those designing ML systems from the ground up. I would highly recommend this book for those looking for clarity in designing ML solutions.
April 26, 2019
Short, extremely rudimentary book which is basically just enough best practice for (I guess) startups to figure out how to develop ML solutions without really understanding them. Overall pretty disappointed by this considering it’s by Andrew Ng whose coursera content I think is excellent, but I suppose it is just a draft after all
March 4, 2019
Brevity is the highest quality of this book. Very sparse on the technical side of machine learning, however, straight to the point. Andrew Ng gives all the important tips on troubleshooting a machine learning system in real life. In summary, a must read, after taking Ng's machine learning MOOC.
December 1, 2019
Read the draft sent by Andrew Ng, hopefully he will complete the book very soon. I learnt a lot from his Stanford videos lectures and this book gives quite broad view of ML, can't wait to see complete book.
April 8, 2020
It was OK with nice practical examples and some good principles to inform anyone seeking to implement ML methods. Somehow it felt a little rushed, lacking a good structure. But it was free, so it was an interesting little read.
August 23, 2018
Rating is for the draft version of the book.
October 2, 2018
Straight forward practical advice, not immensely illuminating
October 3, 2018
A very good summary of machine learning best practices from one of the most respected machine learning researchers and instructors in the world.
November 4, 2018
What excels about this book is the fact that it focuses and provides tips on actual ML problems a developer could encounter in a production and live environment.
December 18, 2018
This is probably one of the best introductory to machine learning and deep learning :)
January 18, 2019
Interesting book with fresh look on some problem from one of the guru of deep learning. It's not a detailed tutorial, rather cookbook for peoples who already have some basic experience.
June 12, 2019
Quick and concrete advice on how to carry out a machine learning project, and what systematic approaches one can take to improve one’s models. Helpful!
June 22, 2019
Breaking down the content into small chapters makes it a joy to read.
August 25, 2019
Gran libro para detalles sobre ML, especialmente en lo relacionado con los sets de datos.
Displaying 1 - 30 of 48 reviews