Andrew Ng said in the Coursera ML course that if you know linear regression, logistic regression, advanced optimization tools, and regularization, you may know more ML than many engineers using ML in Silicon Valley. That's right?

Updated on : December 6, 2021 by Leo Williamson



Andrew Ng said in the Coursera ML course that if you know linear regression, logistic regression, advanced optimization tools, and regularization, you may know more ML than many engineers using ML in Silicon Valley. That's right?

I love Andrew Ng. His courses are fascinating. He has a gift for teaching. In addition, I have learned a lot from him, so writing material that criticizes him is not easy.

That said, he is one of the people responsible for putting millions of academics into real-world roles for which they are not qualified.

There has always been a disconnect between academia and the real world. It is very pronounced in this space.

This idea that you could take a "data scientist" and put them in a highly technical role and make them successful has cost companies billions of dollars.

In the real world, machine learning engineers

Keep reading

I love Andrew Ng. His courses are fascinating. He has a gift for teaching. In addition, I have learned a lot from him, so writing material that criticizes him is not easy.

That said, he is one of the people responsible for putting millions of academics into real-world roles for which they are not qualified.

There has always been a disconnect between academia and the real world. It is very pronounced in this space.

This idea that you could take a "data scientist" and put them in a highly technical role and make them successful has cost companies billions of dollars.

In the real world, machine learning engineers are highly skilled programmers with tons of data knowledge. We are not pontificators, we are doers.

This is why the "data science" role has dropped five years in a row in job boards and the main job is now the machine learning engineer. I am one of the few who has said this was going to happen and I said it five years ago. Everything is documented in my Quora posts.

When you go to work for a company and that company pays you 250K or more, you better be able to show that your models add critical value to your bottom line. If not, regardless of your title, you won't be around for long.

Another truth that many "data scientists" are coming to terms with is the democratization of the model. This simply means that modeling is being automated.

For example, most real-world models are classification and regression on structured data. The king in this space are the gradient drivers. These problems have been democratized. As long as I understand data cleansing, I can pass my dataset to Google's AutoML tables and know that I have the best possible model. You will not exceed the results.

So while it is true that I don't know all the math behind models, I do know how to work the end-to-end machine learning process. In addition, my ability in data cleansing and function engineering has been what has set me apart. Now, you have to be careful how you interpret this. Not knowing the math behind the model doesn't mean you can't apply statistical concepts to your data. This is absolutely necessary, it is called applied statistics and it is completely different from theoretical statistics.

Coursera started the certificate scam. Andrew Ng helped create Coursera. Certificates in any IT vertical are completely worthless. The companies not only do not consider them valid, but they will also filter the resumes with them activated. What you get from my platform is the knowledge you need to get the job done in the real world, not fake posters.

Now, I have learned that I cannot tell you this and make you believe it. You have to experience it. For many of you, it will be a difficult and possibly expensive lesson. If you ever take on a role in applied machine learning, you will learn what it really entails.

EDIT: Added 6/22/2021 - Yes Andrew Ng has read this. No, I haven't heard from him, but he surprised me recently with this quote.

"The model and code for many applications are basically a problem solved," says Ng. "Now that the models have advanced to some degree, we have to make the data work as well."

EDIT: Added 5/23/2021

Here's a fantastic answer that I thought of including. It is the most honest information you will get on Quora.

I have interviewed dozens of candidates for deep learning and deep learning positions at top tech companies. I found that 90% of the candidates lack basic statistical stripping of ML models and techniques, including linear and logistic regression. If you understand these topics deeply, you will know more than they do in that area. But if you only get the basic concept and lose statistical and mathematical understanding, you will be one of them. What Andrew was referring to was "if you know these issues thoroughly, you are ahead of them."

However, even in that case, you know more than they do only in the

Keep reading

I have interviewed dozens of candidates for deep learning and deep learning positions at top tech companies. I found that 90% of the candidates lack basic statistical stripping of ML models and techniques, including linear and logistic regression. If you understand these topics deeply, you will know more than they do in that area. But if you only get the basic concept and lose statistical and mathematical understanding, you will be one of them. What Andrew was referring to was "if you know these issues thoroughly, you are ahead of them."

However, even in that case, you know more than they just on these specific topics. We use many more techniques in the day-to-day of technology companies. Plus, you learn a lot during implementation and handling of real-world problems. Listening to machine learning courses and solving engineering problems is not enough. That 90% is ahead of you in experience. However, at the end of the day, whoever understands best will win the game.

P / S: I started learning ML with Andrew's courses years ago. His courses seemed the best to start with and I learned a lot from him. Anyway, I think his reputation partly resembles that of a celebrity. So, I don't think you should always assume that he is right in whatever he is saying.

Well, if it's true, you're in luck. I think Harvard Business Review predicted that there will be a shortage of about 200,000 data scientists by 2018.

The Prof. Ng class is a good first choice. Everything from linear models to neural networks is really just the subject of neural networks. After that, do unsupervised learning with a couple of classic techniques.

But unsupervised learning is also being done with deep learning / neural network approaches.

It could have changed its class by adding a bit more stochastic gradient descent and mini-lots. You don't just use them for functions with l

Keep reading

Well, if it's true, you're in luck. I think Harvard Business Review predicted that there will be a shortage of about 200,000 data scientists by 2018.

The Prof. Ng class is a good first choice. Everything from linear models to neural networks is really just the subject of neural networks. After that, do unsupervised learning with a couple of classic techniques.

But unsupervised learning is also being done with deep learning / neural network approaches.

It could have changed its class by adding a bit more stochastic gradient descent and mini-lots. Not only do you use them for functions with local minima, but they also work best on functions with a single minima.

Also, with temporal data, you need recurring neural networks. This then brings up the "disappearing gradient problem" and how to deal with it. So this could easily happen after you finished your lectures on neural networks.

It's really about backward propagation and selection functions.

Then the convolutional networks would follow and those are the interfaces that enter the neural networks. They are used in unsupervised learning in many areas and keep all records so far in classifying images, text, speech, language, audio, etc. Extremely basic functions are started with which these virtual networks will find the higher order functions for you. They've been kicking butt lately.

I have taken the same course and I remember the quote. I think it's probably accurate. Most engineers using ML do not need a high level of understanding. Many may be using existing libraries and not understanding what goes on under the hood. I've also seen a lot of companies claim to do amazing things with ML, but when I look at the technical descriptions they provide, it's pretty ... mediocre. They will claim to use a specific algorithm as if it were something special, but it is quite common. Someone who knows nothing about ML may be impressed when you say you are using a k-means cl

Keep reading

I have taken the same course and I remember the quote. I think it's probably accurate. Most engineers using ML do not need a high level of understanding. Many may be using existing libraries and not understanding what goes on under the hood. I've also seen a lot of companies claim to do amazing things with ML, but when I look at the technical descriptions they provide, it's pretty ... mediocre. They will claim to use a specific algorithm as if it were something special, but it is quite common. Someone who does not know anything about ML may be impressed when you say that you are using a k-means clustering algorithm, but it is a really basic technique.
There are much deeper and more interesting topics in ML than those listed, so you're not saying this is the bulk of machine learning. Only most people use the simplest varieties.

Probable. Many software engineers don't bother learning anything but code, resulting in sloppy applications and violating key data assumptions. It's shocking that the people in charge of machine learning don't understand generalized linear models and their assumptions. And it is something that will cause regulatory problems in the future for companies under government regulation for products and services.

I would add tree-based methods to the list to find out, as they are pretty useful these days too. Clustering is a nice addition to unsupervised learning.

Many of the basic principles behind machine learning are independent of the particular model, whether you are working with simple linear regression or some fancy new neural architecture that just appeared. For real-world applications, you also need key data cleaning / pipeline skills that again tend to work the same way regardless of the model used in the end (and this cleaning / pipeline often takes much more time than actual modeling).

So it makes sense to first learn the basics and get to where end-to-end pipelines can be worked on using simple models because once you can do that, you don't need to.

Keep reading

Many of the basic principles behind machine learning are independent of the particular model, whether you are working with simple linear regression or some fancy new neural architecture that just appeared. For real-world applications, you also need key data cleaning / pipeline skills that again tend to work the same way regardless of the model used in the end (and this cleaning / pipeline often takes much more time than actual modeling).

So it makes sense to first learn the basics and get to where you can work pipes end-to-end using simple models because once you can do that, it's not that difficult to swap them out for more elaborate ones. (With that said, you still need to have a good understanding of each model you want to use to understand the tradeoffs between using different models for different situations.)

Many open source tools do things like use bubble sorting or the trapezoidal rule for auc and are either not computed or very slow at a magnitude of 10 to 100x. This also applies to tools like tensorflow, but figuring it out is a bit more difficult. Silicon Valley also has schools that are not geared towards math or CS like Stanford or Berkeley are, but there still need to be jobs for them.

I think that statement was made a couple of years ago when you started your online course, which was probably true during that time. Now, I don't think that's true since you have Deep Learning in the mix. But you never know. I've interviewed hundreds of people on Machine Learning and Deep Learning, and most of them fail spectacularly.

Very likely.

Considering his experience working with Silicon Valley engineers and his integrity towards the ML community, I would agree with this, although I do not recall this quote when I completed his courses.

Not really. What NG wants to tell you is that regression is very useful in everyday work.

Andrew Ng is a super smart individual and I really like the boy.

However, I will be brutally honest about my initial observation of the first 1.5 weeks, which I passed yesterday with great anticipation and really enjoyed (still do!), What I experienced.

In reality, this may have nothing to do with your capabilities or intentions, but rather ("the dilemma") stems from the latest (near-insane) trend of packaging a deep learning course into a MOOC and trying to teach you to the people everything in a pile of walnut shells.

I'll get to that in a minute, but first my quick analysis of who this D is

Keep reading

Andrew Ng is a super smart individual and I really like the boy.

However, I will be brutally honest about my initial observation of the first 1.5 weeks, which I passed yesterday with great anticipation and really enjoyed (still do!), What I experienced.

In reality, this may have nothing to do with your capabilities or intentions, but rather ("the dilemma") stems from the latest (near-insane) trend of packaging a deep learning course into a MOOC and trying to teach you to the people everything in a pile of walnut shells.

I'll get to that in a minute, but first my quick analysis of who this Deep Learning course / specialization may or may not be for.

So who could this course be for?

  1. Movers / Players in the game? The course will not reveal new things if you are looking for additional information and have been in the industry for a while. In fairness to Ng, he advises moving fast for those who have been on a fast track in certain modules like calculus and others. If you are looking to repeat, it is up to you.
  2. Absolute beginners? This isn't for absolute beginners either; even though Andrew tried to simplify it pretty well. In some cases, you have a hard time finding the right approach, course, or technique to explain difficult things with ease. But I don't blame him because it seems like they want to capture a large audience (it seems that way).
  3. Army of employees in the company? Is it for companies that want to transform their entire workforce into tribes of deeper learners? Andrew is subtly targeting companies and maybe there are some hopefuls who want to put their workforce in some kind of exercise and eat up the HR / Education budget, but I don't think that's the right thing to do either (unless you're in a massive cat business)
  4. Lastly, this whole certification thing: Really, you have to ask yourself what or who are you doing this for. Many people have asked me if this can help them get a job at Google, Facebook, Amazon, Netflix. I think we may be (subconsciously) fooling a lot of people into failure, as in addition to a bunch of high-tech companies doing interesting R&D and product work, a lot of this is not so much. rampant as it could be done. to believe. But that's for part 2 (see below).

My personal conclusion: However, I really enjoyed the Geoffery Hinton interview at the end of the first week's session, especially on the "theory of mind and thoughts", which is very similar to what I started working on in 97 in book form. But i stopped because people thought i was crazy or dumb or both!

Glad to see that there is (at least) another on this opposite path / thought. I have resumed my own project after 20 years and I am very happy about it!

But I'm rambling for what I must apologize for!

Note: I will complete this course and (maybe) end up with a specialization. Learning from / with Andrew, even if it is an online / distance education, is a pleasure and time well spent.

Part 2: MOOC: are we missing something here?

A much bigger problem is emerging as more platforms like Udacity, Courera, Edx compete (mainly with each other) for students online and also for their wallets.

Articles mourn over the failure or collapse of MOOCs and indeed, there was an article in the Dutch Financial Times today about the huge churn rate as millions navigate from one platform to another to (hopefully) learn and eventually get certified.

The problem / risks we are actually facing is simply the following:

  1. Universities simply cannot keep up with the change and still teach Java with books that are so old that they belong in a museum. However, it turns out that they are accredited, so leaving college is still important. (Phew!)
  2. MOOCs, on the other hand, like Udacity, Coursera, Edx, no matter how cool and packed with top-notch training for jobs today and tomorrow, are not accredited / recognized by the world outside of Silicon Valley and by companies / businesses that They are not the ones I mentioned above. I wish they were more important.
  3. More serious problem: As students, we all need an ecosystem that we can physically go to, in contrast to the scenario of “sitting alone behind your desk” in solitude battling an unknown monster. You would like to do it together with your fellow students, build relationships, and do more things like brainstorm and build companies and the like. Universities still provide that (physical infrastructure, etc.)
  4. There's also the potential risk that one or more of these online MOOCs (they're all for profit, you know) will eventually be bought by some high-tech company and lose their independent identity. So what?

The solution?

Whether it's studying solar power technology at Delft University through Edx, learning about new battery technology from some other MOOC, or Deep Learning from Coursera, we will be faced with a dilemma that the online generation will have to grapple with: no it is able to retain all that information later. A few weeks have passed. Things just slip out of our heads a lot faster than it used to!

I think this is the most concerning aspect for MOOCs (in their current form in general) and it is also the most fascinating deep learning / AI problem in itself that needs to be fixed.

A hybrid combination of / continuous / for life / with mentors / online / offline could / should offer comfort to this current problem.

Something that offers a solution rich in experiences but at the same time personalized so that you really learn something new!

I would really love for this to happen, but I don't know how.

NOTE: I joined Coursera (as of October 2017) as a mentor for this specialization to do my part well and see if I can make a difference in the students' learning experience.

Other Guides:


GET SPECIAL OFFER FROM OUR PARTNER.