Comment by dalbasal | Hacker News Reader

Comment by dalbasal | Hacker News Reader

Parent Full thread

At least from the 1st few comments, it seems that people want to point out the ploy to get you using MS' products/tools rather than give you a wide, general education.

So... I don't mind this. Learning via a toolset is an OK approach, in a lot of cases. I personally like it. Realistically, a big part of basic competency is familiarity with a functional toolset anyway. Some people will expand their toolsets later.

Also, if MS succeed in providing good learning tools as a way of creating demand for their services, this is cool too. Ideally AWS and others will continue to do the same, and there'll be an incentive to create and maintain free courses.

If you want a more general/abstract course, these exist too. This just isn't it.

Based on a quick glance I'd say the course wants to get you to a point where you connect to data, do a little cleanup coding, use it to build a model & then turn it all into a web service. It probably has a little bit of stats & ML "theory," but not a lot. This is useful if its useful to you. If you're mostly interested in theory, this just isn't for you. Try the big coursera course, maybe.

Speaking as a CS educator, I agree, with a caveat:

In my free time, I took the big Coursera course, and it is (was?) also tied into a proprietary toolset, GraphLab. They have "you can also do this in pandas" instructions on lab assignments, but the instruction and tools are aligned to GraphLab. I imagine the vast majority of students will not even consider using pandas.

Graphlab Create has since been open-sourced (as Turi Create) by its new owner, Apple. I didn't have a problem with the setup back then.

Stressing over toolsets is like stressing over your first programming language. It's counterproductive. You will do a lot better by finding a source of instruction you trust and a toolset you can work with, and then actually making something.

Quality instruction vastly outweighs the tradeoffs with learning a new language/api/etc.

From having skimmed the Microsoft course's offering, it looks like it would prepare a good student to understand and apply the algorithms to make models, but not to implement the algorithms themselves (which the Coursera course does do).

Which is fine! I could tell you how to implement elliptic curve cryptography and prove most of the mathematical underpinnings, but I have no intention of ever writing an implementation myself.

fast.ai’s course looks great, and they even have a numerical linear algebra course.

For a theoretical background, I particularly like the Shai-Shalev Schwartz book, though it’s not about deep learning. I’m told the Goodfellow book is good only in the sense that a better book simply doesn’t exist.

If you are looking for profound theoretical insights in Deep Learning, you will be disappointed, but amazed with classical ML. If you are looking for practical capabilities, you'd be amazed by Deep Learning but frustrated with classical ML. Deep Learning is fairly trivial math-wise (unless you look carefully into non-linear optimization it uses, which is as brutal as any other advanced part of math).

Agreed. I have used some of the Azure ML services a bit and its really easy to understand concepts and visualize models via a designer, like a workflow engine. Pretty neat. I havent used it for anything professional yet though.

Reply on news.ycombinator.com