Selected from Medium
Author: Arvind N
Translated by Machine Heart
Contributors: Lu Xue, Li Zenan
On August 8, Andrew Ng officially launched Deeplearning.ai—a series of deep learning courses based on Coursera, aiming to spread foundational knowledge of artificial intelligence to more people. A week later, many have completed the first three courses that are currently open. Who are these new courses suitable for, and can they compare with the classic “Machine Learning” course? Let’s take a look at this preview experience.
Between full-time work and family chores, many people hope to use their spare time to learn about cognitive science and artificial intelligence. If a great course suddenly appears, everything becomes simpler.
The deeplearning.ai course recently launched by Andrew Ng is just that.
The course was released on Coursera, and I immediately registered and spent four evenings watching lectures, taking exams, completing programming assignments, and passing the course.
Deep learning practitioners and machine learning engineers often spend a lot of time on abstractions like Keras and TensorFlow. But taking some time to deeply understand the essence of learning algorithms and manually writing backpropagation can be a very meaningful process. It is not only interesting but also allows you to learn a lot!
Contents Included in Deeplearning.ai
As a new attempt by deep learning pioneer Andrew Ng, Deeplearning.ai is a course system that teaches you the principles of neural networks from the ground up, while being easy to understand, at a beginner to intermediate level.
This is the classic Andrew Ng style, where knowledge is presented through carefully selected courses, appropriately timed videos, and precisely set information blocks. Andrew starts with the information omitted from his famous “Machine Learning” course, introducing neural networks from the perspective of a single neuron (logistic regression), and then gradually increasing complexity—adding more neurons and more layers. By the end of the four-week course (the first course), students will learn all the core knowledge needed to build complex neural networks, such as loss functions, gradient descent, and vectorized parallel Python (numpy) implementations.
For the mathematics and programming concepts behind deep learning, Andrew also provides a step-by-step patient explanation, ensuring that students can maintain interest in both mathematics and code.
Course Materials and Tools
Video Lectures
The lectures use slide presentations, accompanied by Andrew Ng’s own notes. This seems to be a method that keeps people focused at all times. I find it best to set the video speed to 1.25x or 1.5x.
Testing Tools
Tests are scheduled after each lecture video in the form of multiple-choice questions. If you watch the video completely, these questions should be easy to answer. You can also retake the tests multiple times, and the system will keep your highest score.
Jupyter Notebook Programming Assignments
The programming assignments need to be done using Jupyter Notebook, which is a powerful web-based application.
The assignments have a very good guiding sequence structure, and you only need to write two to three lines of code in each blank. If you understand concepts like vectorization, you can also complete most programming problems in one line!
After completing the assignment, you need to click the button to submit your code, and the automatic grading system will return your score in a few minutes. Some assignments have time limits—like only three submissions within eight hours.
Jupyter Notebook is well-designed, with no bugs; it feels like a polished industrial product.
Who the Course is Suitable For
Anyone interested in understanding what neural networks are and how they work; and those who want to build neural network tools and are interested in tools that turn ideas into reality are suitable for this course.
Not good at math? Don’t worry, Andrew explains all the necessary calculus issues and explains the derivations in each case, so you can focus on building neural networks and implementing your ideas through code.
If your programming isn’t good enough, there are assignments in the course that teach you how to use numpy. But I still recommend you learn the basics of Python on Codecademy beforehand: https://www.codecademy.com/learn/learn-python.
Difference Between This Course and Jeremy Howard’s Fast.ai Course
Let me explain with an analogy: imagine you are learning to drive.
Jeremy’s FAST.AI course starts teaching you from the driver’s seat. He teaches you how to turn the steering wheel, hit the brakes, and press the gas pedal, etc. Then he explains how the car works in detail: why turning the steering wheel makes the car turn, why hitting the brakes slows down and stops the car, etc. He takes you deeper into understanding the internal workings of the car, and by the end of the course, you understand how an internal combustion engine works, the design principles of the fuel tank, etc. The course is designed to teach you how to drive. You can choose to stop learning at any point after you learn to drive if you feel it is unnecessary to learn how to build or repair a car.
Andrew Ng’s deep learning course covers all of the above content, but in the completely opposite order. He first teaches you about the internal combustion engine! Then he gradually increases the level of abstraction, and by the end of the course, your driving skills are like an F1 driver!
The Fast AI course mainly teaches you how to drive, while Andrew Ng’s course primarily teaches you the engineering principles behind the car.
How to Study This Course
If you have no prior knowledge of machine learning, then don’t take this course yet. It is best to first learn the machine learning course previously published by Andrew Ng on Coursera (https://www.coursera.org/learn/machine-learning).
After completing that course, take the first part of Jeremy Howard’s excellent deep learning course (http://course.fast.ai). Jeremy teaches deep learning from the top down, which is necessary for beginners.
If you can successfully build deep neural networks, then you can study this new deeplearning.ai course (https://www.deeplearning.ai). If you have any misunderstandings about the underlying details and concepts, this course will clarify your doubts.
What I Like About This Course
-
The facts are presented clearly—periodically eliminating any uncertainty and ambiguity.
-
Andrew emphasizes the engineering aspects of deep learning, providing numerous practical tips that save time and money. As the lead engineer in my engineering team, I benefited greatly from the third course of this program.
-
The handling of terminology is excellent. Andrew believes that the empirical process is a trial-and-error process. He candidly states the realities of designing and training deep networks. At times, I felt he might think deep learning is just curve-fitting.
-
He dispels all the hype surrounding deep learning and artificial intelligence. Andrew provides restrained and serious commentary on the media’s hype about AI, and by the end of the course, you can be clear: deep learning is not the Terminator.
-
Good boilerplate code is ready to use!
-
Excellent course structure.
-
Consistent and useful notation. Andrew attempts to build a new naming convention for neural networks, and I think he succeeded.
-
Andrew’s unique teaching style continues the style of his previous machine learning course. I can feel the same excitement as when I learned deep learning in 2013.
-
The interviews with prominent figures in deep learning are refreshing. Hearing the personal stories of these pioneers is very inspiring and interesting.
What I Think This Course Lacks
I wish Andrew would say “concretely” more often!
Additional Insights I Gained from This Course
1. Deep learning is not simple. You need to spend a lot of time and a lot of hard work to “learn” the concepts and get the models to work properly. Andrew recently wrote a Quora answer (https://www.quora.com/How-should-you-start-a-career-in-Machine-Learning/answer/Andrew-Ng?srid=TS2A) that resonated strongly with me.
2. Good tools are very important and will help you accelerate your learning. Seeing Andrew use a digital pen while teaching made me buy one too, and it helps me work more efficiently.
The black ink is Andrew’s handwriting, and the other colors are mine.
3. I recommend that everyone learn the Fast.ai course first for a psychological reason. Once you find your passion, you can learn freely.
4. Every time you score full marks, you get a dopamine rush.
5. Don’t be intimidated by deep learning terminology (hyperparameters = settings, architecture/topology = style, etc.) or mathematical symbols. If you try boldly and listen carefully, Andrew will show you why these symbols and notations are so useful. They will soon become your favorite tools!
Some symbols that look scary. As you start watching the course videos, these symbols will become easy to understand.
Conclusion
1. Everyone starts learning or working in this field as a beginner. If you are a beginner in deep learning, it is natural to be intimidated by these terms and concepts. But please don’t give up. You may be attracted to this field and find your mission. Trust your intuition, stay focused, and you can achieve success! Even Andrew Ng had to learn linear algebra, right? He wasn’t born with this knowledge.
2. Although this is a very good resource, it is not the only deep learning course in the world. Many generous teachers, like Salman Khan, Jeremy Howard, Sebastian Thrun, Geoff Hinton, and Andrew Ng, share their knowledge for free online. I wasn’t very lucky; to find a job and support my family, I didn’t pursue a PhD, but that doesn’t mean I will stop learning. Thanks to the democratization of knowledge, I have the opportunity to set my own learning agenda and learn from the courses of the people I admire the most: programming (Gerald Sussman), linear algebra (Gilbert Strang), AI (Marvin Minsky), philosophy (Daniel Dennett), psychology (Jean Piaget), physics (Hans Bethe).
3. Most applications of deep learning are indeed rigorous engineering problems. Professor Andrew provides very interesting explanations in the third course (my favorite among the three courses). The mindset needed to solve problems with deep learning is the same as that needed to solve any complex engineering problem. Everything you need to understand has been clearly written by Claude Shannon decades ago (https://medium.com/the-mission/a-genius-explains-how-to-be-creative-claude-shannons-long-lost-1952-speech-fbbcb2ebe07f).
Original article link: https://medium.com/towards-data-science/thoughts-after-taking-the-deeplearning-ai-courses-8568f132153
This article is translated by Machine Heart, please contact this public account for authorization.
✄————————————————
Join Machine Heart (Full-time Reporter/Intern): [email protected]
Submissions or Reporting Requests: [email protected]
Advertising & Business Cooperation: [email protected]