I thought it was useful but awfully low level. For example I hope to never, ever implement backpropagation again; I'm going to use whatever code is in TensorFlow or PyTorch or whatever. But as a student I'm glad I did implement it myself, once, so I understand what is going on. More broadly it demystifies the black box of machine learning methods and you can see it for the giant pile of statistical categorizing functions that it is.
The most practical takeaway I got from Ng's course was the dangers of under and overfitting your data and techniques for detecting when you make that mistake.
I still remember a talk by a woman from Google at a fairly long ago now O'Reilly conference (R.I.P). Part of what she discussed was Research AI vs. Applied AI. The gist of it was that a lot of the things in university course, graduate programs, etc. are tilted towards Research AI and you can get away without a lot of that stuff by using pre-built tooling for practical machine learning applications.
Of course, you want to have some understanding of what's going on under the covers but, for a lot of people, starting from first principles is quite hard and isn't really necessary.
Not the course I took. It relies on basic linear algebra like matrix multiplication. You can probably get through it with just coding and not understanding the math but it wouldn't be much fun.
I took this course as a defensive mechanism against BS at work, especially when the consulting Data Scientists were around. In that sense it's super practical.
ML is dominated by gigantic datasets and massive computing powers, something individuals will not have a lot of.
It is unlikely that you could build a major product with it, but it could tech you neat tricks to speed up some parts of work. Also, similar to cs101, it is a necessary first step towards a career in ML. So might as well do it.
I know a bunch of business analysts and data analysts who have gotten a job based on what they learnt in this course. Ofc, they also got some stem degre alongside it, but this course made a difference.
In 2012 I did Andrew's original machine learning course, and implemented a bespoke OCR engine for iOS, which was released in a banking app for scanning utility bills. Back then deep learning was just taking up, so I did my own backprop training in Matlab based on Andrew's code as well. It was a pretty fun end-to-end experience, much better than just throwing stuff at tensorflow like we do nowadays.
Do you thing now days Deep learning does not requires much math? If yes, to what extend of knowing math is enough to be truly good deep learning specialist? By deep learning specialist I mean the person who is building a commercial software that uses deep learning but not tools for deep learning.
The things I learned here helped me gain a solid foundation, which, in turn helped me learn Deep Learning.
And Deep Learning feeds me now.
The good thing about this course is that it is not Math-shy. It is not rigorous in terms of Math, like there are no proofs and so on. But Math is omnipresent here.
Andrew Ng's MOOC is among the best game in town. Ng is among the best teachers I have ever seen.
No. The ugly truth is that these courses will be useless to 99% of the people. Machine learning is dominated by big corporations with gigantic amounts of data and processing power. If you want to work in one of them or create competing ML companies you need pedigree (a PhD from a well know university), and those guys arent taking courses with fake credentials.
You could use ML in your job/company but then you dont need this course, you just use a ML product.
See this course as a hobby thing, or if you are in HS and want to start preparing for college, otherwise there are better uses of your time.
There's a lot of ML happening outside of big corporations, which you can confirm by just searching 'machine learning' on any job site. While it's true that often you can use ready-made ML solutions, you often will benefit from additional knowledge for improving or adjusting them for your company's specific problem and while interviewing you will often be asked the kind of questions those courses cover.
You can get almost unlimited GPU time on Google Colab for $50 a month. I don't know why or how they pay for this, but it does bring "real research" into the reach of individuals.
You can get more processing power and true unlimited time with any semi-competent graphic card (probably costing less than 1 year of Colab Pro+). Pro+ is a scam, you are not told what kind of instances you will be running at, and you dont have any guaranteed continuous running time. And even if you were given full 24/7 access to a top of the line card that would be like 0.001% of the power used to train any big modern ML model.
For example, Google Vision API can do some out-of-the-box classification on arbitrary images with no training needed. Covers super common cases such as explicit content detection and object detection.
There are more customisable products within Google where you can provide training examples and labels using a UI (AutoML I think it's called). The result is an endpoint you can use to do inference, based on the model created behind the scenes.
I just mention these examples because I've spent a little time researching them at top-level.
The most practical takeaway I got from Ng's course was the dangers of under and overfitting your data and techniques for detecting when you make that mistake.