Chevron Left
Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization に戻る による Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization の受講者のレビューおよびフィードバック



In the second course of the Deep Learning Specialization, you will open the deep learning black box to understand the processes that drive performance and generate good results systematically. By the end, you will learn the best practices to train and develop test sets and analyze bias/variance for building deep learning applications; be able to use standard neural network techniques such as initialization, L2 and dropout regularization, hyperparameter tuning, batch normalization, and gradient checking; implement and apply a variety of optimization algorithms, such as mini-batch gradient descent, Momentum, RMSprop and Adam, and check for their convergence; and implement a neural network in TensorFlow. The Deep Learning Specialization is our foundational program that will help you understand the capabilities, challenges, and consequences of deep learning and prepare you to participate in the development of leading-edge AI technology. It provides a pathway for you to gain the knowledge and skills to apply machine learning to your work, level up your technical career, and take the definitive step in the world of AI....




Fantastic course and although it guides you through the course (and may feel less challenging to some) it provides all the building blocks for you to latter apply them to your own interesting project.



I enjoyed it, it is really helpful, id like to have the oportunity to implement all these deeply in a real example.

the only thing i didn't have completely clear is the barch norm, it is so confuse


Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization: 51 - 75 / 7,029 レビュー

by Alec T


Outside of an academic space, I think this course is one of the best places to be introduced to deep learning. It think it does a tremendous job of orienting you around issues relating to implemention of neural networks for deep learning. The information if transferred in an efficient and concise manner - the videos are created in a way that are not too long or too dense in any one sitting to be overwhelmed. I do think this course is best supplemented or partnered with a course that requires more implementation and generation of neural networks. The activities can be performed swiftly if you have been paying attention to the lectures, which may leave some wanting slighty more from an application perspective. I can however admit this is just the 2nd course of 5 in the specialization so take this information with a grain of salt.

by Johan D R P


I found this course pretty useful to understand a large set of options to explore in a Neural Network (and its inputs) in order to improve its performance. It shows the mathemical fundamentals behind the concepts, without going too deep to confuse a person without advanced calculus knowledge, like me.

However, I would like that the following changes were made:

Update last week lab to Tensorflow 2.0. This framework update seems to be more beginner friendly, because its interaction with Python functions (no need for sessions). Also, a lot of things shown in the lab are deprecated.

Make a lab for hyperparameter exploration. While this task can take a relatively long time, maybe it would be feasible to explore hyperparameters over a simple model. I felt that the course needed more hands on in the part of exploration.

by Anirudh K


This is a really informative course and really crucial if you are planning to do a personal project or even prepare for interviews. It equips you with all the tools to get started with actually start implementing Neural Networks for a problem by 1) Teaching how to prepare data sets 2) Regularization/dropout to increase accuracy on test set, 3) Set up your optimization problem 4) Teach different Optimization Algos 5) Teach Hyperparameter tuning and the order of importance of different hyperparameters 6) batch Normalization and lastly Tensorflow. Andrew NG is truly a master in teaching concepts in an approachable and intuitive way. I believe the course can be made even better by adding Keras to the programming frameworks module along with more videos and programming exercises for data pre processing.

by Valery R


T​he content of the course is great, but I didn't like at all the final programming assignment on TensorFlow. I'm not a fan of all those third party libraries, or software (not sure how this should be called) which add one more layer between the user and the machine. We call functions without really knowing what we are doing. We just do what the assignment tells us to do. That's relatively easy, but it the assignment could be done without understanding the content of the course. All we have to is simply follow the instructions. We end up with a 100% grade, and we almost have no clue about what we did. I prefer much better to program my own implementation of the neural network in C++ (or Python since this is fashionable), which really forces me to understand what I am doing.

by Nigel S


It explains a bunch of complicated maths and methods in a way that is at least comprehensible by mere mortals, though not necessarily easy. Put another way, if this course doesn't enable you to understand how to tune and optimise deep neural networks, then you probably never will.

The content taught in this course is really valuable because it explains a lot of what is going on behind the scenes in the existing Deep Learning Frameworks like Tensorflow, Keras, etc, and enables you to be a lot more competent and confident in producing effective models in a time-efficient way, than if you didn't have this knowledge.

It also seems to have been built by peopel who not only know the material intimately, but who recognise that many of the learners are very time-poor.

by Aakash S


Just as I started this course, a person building a DNN using Tensorflow and Keras reached out to me for help. He had spent a day struggling to figure out how to improve the performance of his model to improve the accuracy for training and validation (aka dev) dataset, and then figure out why his model was so inaccurate for test set, even as his model was very accurate for the dev set. After going through just the first week lessons of this course, I was able to suggest to him what approaches to take to optmize hyperparameters of his model to get the desired accuracy for all datasets (tarining, dev and test). It seldom happens that one gathers enough insight after just a week of lessons to apply it towards solving real world problems!



This was amazing stuff from the team. Very good to get going on Deep Learning. I have some comments on the course curriculum, regarding the optimization techniques(OTs). The theory presented here is good, but as a learner seeing it for the first time, I didn't get a deep intuition on why we are doing a particular OT and why we need different OTs. So I learned this part from the NPTEL course on Deep Learning by Mitesh Khapra, where he first mentions the reason behind developing an OT, and then he deep dives into the implementation of the technique. In a great way, he developed the concepts on Momentum, NAG, Adagrad, RMS Prop, Adam progressively. Probably an extra video on the analogies behind the OTs can be added on coursera:)

by Taylor B


I took the Machine Learning Course from Stanford with Andrew Ng a few years ago and enjoyed it but I was also somewhat overwhelmed by the math. In contrast, this is my second course in the deep learning specialization and I feel like so far the courses have struck a good balance, introducing core concepts and derivations for things but also making sure I get guided practice along the way, and also not moving straight to frameworks but having students code more or less from scratch first. I'll probably need some practice on kaggle or other datasets as well as reference to a few other learning materials to feel like a strong practitioner, but this gives the tools to make that possible and I'm very satisfied with this result.

by Jorge L


All the courses in the Deep Learning Specialization are very good and met my expectations. I was guided through the nitty-gritties of neural networks, fortunately with a strong emphasis on Computer Vision (my area), deep diving in coherent coding exercises. Prof Andrew, as always, managed to connect the points between theory and practice, recollecting the concepts treated in past lectures, while showing how Tensorflow operates and how to use it. If you ask me, I'd say that the slides of the Machine Learning course used to be better than the slides for the 4 courses in this specialization, in the sense of being useful as studying guide for the future. The current slides only make sense to those who went through the course.

by Luca C


Knowing this makes the difference. How do you evolve from being a monkey behind a keyboard knowing how to tensorflow a NN to homo sapiens? The concepts provided in this course will make the job.

pros: + workflow to address and optimize your supervised learning problems

+ wide and easy-to-get overview on most essential concepts

+ improves your understanding of NN; those who are already familiar with these concepts might still benefit from this clear and insightfull presentation

cons: - programming assignment will not suffices to give you a sufficient knowledge of tensorflow to make your own applications, you should integrate a bit. (However, mastering tensorflow is not the intention of the assignment).

by Maxime


Très bon complément de la partie 1 avec systématiquement une présentation intuitive puis technique (pour programmer à partir de 0) des concepts ..Le contenu est très dense, je conseille donc de prendre des notes et d'essayer de refaire quelque démonstration! Je conseille de bien lire tous les codes en détail ligne à ligne.

Par contre l'introduction à tensor flow n'est pas très bien expliquée selon moi (seul petit point faible) mais ce n'est pas l'objet de cette partie je pense.

Attention la deuxième partie ne comporte pas de sous titre français. Mais finalement en ayant un niveau moyen en anglais, j'ai facilement suivi avec les sous titre anglais car il y a souvent des schéma et le même vocabulaire revient souvent.

by Baohe Z


Good pace for beginner as the last one. With step by step teaching us a lot of useful skills to train our model much faster, Andrew starts to put more attention on practical field, and rather than giving us many equations, he as before likes to use some vivid examples for giving us an intuition, which I think is very helpful to understand those scientific words of computer science. But it doesn't mean, that this course is perfect, even I gave a full point to it. The subtitles have a lot of mistakes and the audio is also poorly processed. Sometimes, you will hear the same words twice.

But in a word, this is the best course for the beginners and the engineers who are willing to know something about ML and AI.

by Amilkar A H M


I loved it. It showed me practical aspects of machine learning, including how to chose the hyperparameters and how to use tensor flow. My only complain is that I'm not sure how much of this information I will retain given that the practical exercises are guided. They build a lot of the functions for you. Still I'm giving it 5 stars because I have not seen this problem solved so far in any other Coursera course. They need to find a balance between teaching you a lot and making it easy enough for most people to be able to pass the exam and not get stuck in the details. Probably they could offer extra practice automatically graded exercises for those of us who want to make sure we won't forget the material.

by Victoria D


I'd highly recommend this course to any of my colleagues interested in Deep Learning.

It is a great followup to Deep Learning and Neural Networks.

My one 'complaint' is that the mathematical depth is too shallow for someone like me (PhD, Mathematical, Computational and Experimental Physics)

It would be great if there was a course that was targeted to people with advanced STEM degrees, and years and years ( 4 decades in my case) of software engineering experience, where more time was spent on the mathematical framework, and the basic algorithms; that way, I'd have the satisfaction and pleasure of constructing the full algorithm implementations myself.

That being said, once again, Andrew is a great teacher.

by Matthew C


I was very impressed with the quality of Dr.Ng's teaching; simple enough to build confidence in your understanding of the inner workings of neural networks yet thorough enough to prepare you for deeper study (academic or otherwise). For $50 this course is a steal; you could go gather all the information & struggle through it yourself but be prepared to spend a lot of time & effort sifting through mis-information.

After taking the 1st coarse I was impressed; course 2 follows in a similar vain. Some of the courses offered through Coursera are more polished than others; if you're at all curious in deep learning, or even if you've already begun your studies, do NOT miss out on this opportunity.

by Shazib S


Perfect teaching material and syllabus. The lectures are very easy to understand and the way Andrew takes the student through each topic creates a level of understanding I did not have before. Thank you Andrew Ng, Coursera, and

One thing I would like to say is that there are quite a few problems with the sound, especially the trailing emphasis on the 'sss' sounds is very annoying and distracting. Also, there should have been a pointer visible on the screen. As the videos now are, when Andrew says something like "... for this equation here..." I have no idea which equation he is talking about. Kindly resolve this problem. Otherwise this course is 10/10.

by Francis S


Previously, I have taken online classes before in Machine Learning by going the cheap route (Udemy, blogs, youtube) and you get what you pay for. Andrew Ng explains it the most thorough, easiest, and simplest way possible. Presentation material is very understandable. Great class for new machine learning learners. Highly recommend it. The only downside is that the programming exercises are little too easy in my opinion. I feel like the best way to get your hands dirty is to do actual projects (do your own projects). These lectures are good for intuition and background of different types of Neural Network architectures. Other than that, Great material. Thanks Andrew!

by Emilio J


El curso está muy bien impartido por Andrew NG y te permite adquirir muy rápido conocimientos sobre los puntos clave para mejorar el aprendizaje con redes neuronales de una forma genérica. La práctica de programación con la plataforma tensorflow de python es muy valiosa, aunque se hecha de menos una mayor profundidad en el uso de las herramientas disponibles de tensorflow y otras utilidades de python para redes neuronales. El curso utiliza como ejemplos didácticos y prácticas la aplicación de redes neuronales al reconocimiento de imágnes, pero estaría bien ampliar los ejemplos con aplicaciones prácticas a otros campos como puede ser un modelado de un proceso físico.

by utkarsh v


This course has been the game-changer in my understanding of the concepts of hyperparameter tuning and optimization. The conceptual knowledge of various tuning techniques along with the theoretical and practical information about the algorithms like GD with Momentum, RMSprop, and Adams have made me confident.

The introduction to the machine learning framework along with the assignment focusing on Tensorflow has also made me confident to learn more and prepare better projects.

As I have mentioned in my previous course review as well the programming assignments are extremely useful and very much important for the understanding of various deep learning concepts.

by Adam S


Andrew ng is simply the best. He is by far the number one teacher for ML. He explains the materials in such an intuitive way like no one else. I think that for my needs, he introduces just the right amounts of math and practicality.

This course really showed me how gradient descent optimization methods work. From before, I knew about these optimization methods and what they do, but never fully understood them. After taking this course I feel much more confident using them. The transition to tensorflow is done at the perfect time. After writing NNs yourself using numpy (first course in this specialization), you can truly appreciate tensorflow.

by Tianyi X


Thank you Prof. Ng and Cousera for this amazing course! It was totally helping me with my career development and with feeding my interest in Machine Learning science itself!

Upon cancelling my subscription of this course, I would to like to let you know that, it had to happen simply because coursera's webpage wasn't smoothly running with my current network based in China. The browsing and watching experience wasn't very idealized and it sort of discouraged me from keeping using it. If the issue could be solved, I am sure more people in my location will be attracted and encouraged to use Coursera.

Stay safe and I hope you all the best,

by Teguh H


Great in depth explanation from ground up on how to tune parameters. Including many personal experience by Andrew Ng throughout years of experience of handling AI projects. Before going into quick shortcut by using the Tensorflow libraries, it is really useful to know the concepts and intuition on how Deep Learning works from ground up. Also teaches you how to solve with many problems in overfitting, underfit, reading the results. In short, his experience that has seen many researches spent too much time into creating projects, and end up hitting brick walls, is summarised with suggestion on how you can avoid that in your AI project.

by Pantelis D


Another excellent course by professor Andrew Ng, short, on point and clear videos that go into the subject of optimizing Deep Neural Networks.

Like the previous course of the specialization the programming assignments are coded and submitted in the browser using Jupyter notebooks, the coding language used is python and for the math the python library "numpy". In the final week of the 3, an introduction to Tensorflow is made.

It is worth mentioning that some interviews with influential people on the field of DL are included and make the student fall in love with DL even more. Excited to see what's next in this specialization.

by Gerardo M L


The course is amazing, the instructor explains everything with a good level of comprehension. All the covered topics are easy to understand, and the tips given are valuable. The examples given are new including also the information seen in the previous course, so you have a review of parts of the content you have seen. Although he keeps using the cat example, he introduces new other applications that are useful.

I wish that the last assignment were a little bit harder, or that we could use our previous knowledge and complement it with this new, but I suppose that it is this way because of pedagogy and it focus on the topic.

by Ronald A R


I'm so appreciative for the course content and the direction that it has moved me in a deeper understanding of NNs and tuning and optimization. I don't know about other students but I am a serious student that labors to comprehend all of Dr. Ng's videos. The labs are very helpful and gently focus the student on code lines that relate directly to video content. The very last lab compelled me study the TensorFlow function library and find what I needed to get the one line of code correct for computing the cost. This took time but was valuable. I'm now looking forward to more experience with the TensorFlow framework.