I guess he means you don't need advanced knowledge of learning theory in order to apply machine learning algorithms. Not sure if I agree. It's always good to know the limits of the algorithms you are using, but it does require a time investment.
There's a philosophy that, once an algorithm has been discovered, you don't really need to understand it to put it into practice. Why understand how sort algorithms work when "sort" is already built into every library? I'd be curious to see some sort of scientific or formal study on that, but in my experience, the people who view most algorithms as black boxes - they understand more or less what you get out when you put in the right things, but have no interest on how it gets to that point - can produce applications that do work correctly under optimum conditions, but fail in surprising ways when faced with unexpected input.
I'm not the person you asked, but from what I've seen HN tends to have the attitude that you don't need to understand "maths", "algorithms", "statistics", "data structures", or any of that complicated stuff, all you need is to get your hands dirty with code. If you buy that view, then this book is the opposite of what HN usually likes.
Up to a point this applies to the more complex machine learning stuff, but I would argue against people on here not understanding algorithms and data structures. Knowing data structures and algorithms is what makes the difference between a good programmer and an average/bad one.
I'm not the person you asked, but from what I've seen HN tends to have the attitude that you don't need to understand "maths", "algorithms", "statistics", "data structures", or any of that complicated stuff, all you need is to get your hands dirty with code.
I don't get that at all. From what I've seen over the years, there is a clear divide on HN between the people who lean towards the "you don't need any theory at all" people and the "you need to know all the theory to do anything" people, with there being a slight bias towards the latter in general.
Interesting how perspectives differ on this site. I guess it's similar to how some people argue that HN is overrun with government hating libertarians, while other people think the site is overrun by granola eating, tree hugging, hippie communists. :-)
> I don't get that at all. From what I've seen over the years, there is a clear divide on HN between the people who lean towards the "you don't need any theory at all" people and the "you need to know all the theory to do anything" people, with there being a slight bias towards the latter in general.
Haha I think I know what camp you're in, since no one claims you need to know _all_ the theory to do _anything_. ;-)
While we disagree on which group is larger on HN, I'd be open to consider that maybe our difference in perception is entirely explained by selective memory (in the sense that you're more likely to remember when people rubbed you the wrong way than when they agreed with you).
Haha I think I know what camp you're in, since no one claims you need to know _all_ the theory to do _anything_. ;-)
Interestingly enough, I don't really think of myself as falling specifically into either. But I'm a pragmatist, so I lean a bit towards the idea that you can get things done without necessarily always needing a lot of theoretical knowledge. BUT at the same time, I think that having a solid theoretical understanding is almost always desirable, and all other things being equal, I'd rather know the theory than not. :-)
I'd be open to consider that maybe our difference in perception is entirely explained by selective memory (in the sense that you're more likely to remember when people rubbed you the wrong way than when they agreed with you).
Agreed. I would think that is probably a big part of it.
I found this book very useful, personally. It does a great job presenting some of the theoretical foundations in an understandable way (e.g. the PAC-learning framework and VC dimension.) I also particularly liked the explanation on PCA.
This is one of the standard PhD-level textbooks. It's probably a little bit much if you just want to use somebody else's library, but if you're developing your own algorithms or you want to engage seriously with the research literature, you have to know what's in here.
The main other books I'd consider are Foundations of Machine Learning (Rostamizadeh, Talwalkar & Mohri) and Machine Learning: A Probabilistic Perspective (Murphy). There are a few others, but I think you'd need a good reason not to pick one of those three.
Looks coherent, informative and not too indulged with proving things with mathematical formulas. Problem is that the only way I could study it with the meticulousness it requires would be during a course. Studying it at my own leisure is a bit too much of a burden.
However what irks me a little, as it did with my university's introductory machine learning course, is the direct jump to the algorithms while omitting explaining the basis for the underlying cause why it is the right thing to do. Perhaps it's presumed that the reader knows the basic statistics behind the idea that we can make assumptions of the data and predict values due to central limit theorem. Machine learning itself then can be seen as a derivate of that as more free-form, heuristic approach to cases where the central limit theorem can't be applied or is too restrictive.
Am I speaking complete nonsense here? Please those wiser than me, tell me am I right?
Inferential Statistics lets us make estimates about some quantity given a set of assumptions about how a sample was drawn. E.g., you might assume data is distributed normally.
From a theoretical standpoint, Machine Learning basically just explores what happens if you remove that assumption, and try to draw conclusions that apply no matter what the underlying distribution was.
The book does build up that theory. You need to understand the Probably Approximately Correct framework (e.g., how many data points do you need to be 90% sure your algorithm will generalize with less than 5% generalization error) and VC dimension (basically a measure of how much different algorithms overfit.) It doesn't even introduce an algorithm until chapter 8!
VC dimension is sort of the equivalent of O(n) style complexity for machine learning. You can use this framework to estimate the effectiveness of different algorithms from a theoretical standpoint given a certain amount of data. For example, Linear Regression's VC dimension grows as O(n) with the number of features, while Decision Trees grow as O(2^n) with depth. This means that Trees are much more likely to overfit on small data, but more likely to be find complex structure in larger datasets.
I took the author of this book's course during my undergrad and quite enjoyed it. It's a good theoretical introduction to machine learning principles. The video lectures are available here: https://youtu.be/b5NlRg8SjZg
As others have mentioned this is a fairly theoretical take on machine learning which may not be useful if you just want to use a deep learning library. That said, I think there is a lot of value in having a deeper theoretical grasp of a topic even when practicing.
This seems like a great resource for understanding how machine learning algorithms actually work. Anyone reading the book to help them build more accurate models, might also be interested in supplementing it with more research into the importance of feature engineering before training machine learning models.
While feature engineering hasn't been a rigorously studied within the academic literature, this book does have a section on feature generation, which gives some practical tips once your data is in feature matrix form.
For those interested in even more, I work on a python library for automated feature engineering called, Featuretools (https://github.com/featuretools/featuretools/). It can help when your raw data is still too granular for modeling or comprised of multiple tables. We have several demos you can run yourself to apply it to real datasets here: https://www.featuretools.com/demos.
23 comments
[ 3.1 ms ] story [ 69.0 ms ] threadI don't get that at all. From what I've seen over the years, there is a clear divide on HN between the people who lean towards the "you don't need any theory at all" people and the "you need to know all the theory to do anything" people, with there being a slight bias towards the latter in general.
Interesting how perspectives differ on this site. I guess it's similar to how some people argue that HN is overrun with government hating libertarians, while other people think the site is overrun by granola eating, tree hugging, hippie communists. :-)
Haha I think I know what camp you're in, since no one claims you need to know _all_ the theory to do _anything_. ;-)
While we disagree on which group is larger on HN, I'd be open to consider that maybe our difference in perception is entirely explained by selective memory (in the sense that you're more likely to remember when people rubbed you the wrong way than when they agreed with you).
Interestingly enough, I don't really think of myself as falling specifically into either. But I'm a pragmatist, so I lean a bit towards the idea that you can get things done without necessarily always needing a lot of theoretical knowledge. BUT at the same time, I think that having a solid theoretical understanding is almost always desirable, and all other things being equal, I'd rather know the theory than not. :-)
I'd be open to consider that maybe our difference in perception is entirely explained by selective memory (in the sense that you're more likely to remember when people rubbed you the wrong way than when they agreed with you).
Agreed. I would think that is probably a big part of it.
The main other books I'd consider are Foundations of Machine Learning (Rostamizadeh, Talwalkar & Mohri) and Machine Learning: A Probabilistic Perspective (Murphy). There are a few others, but I think you'd need a good reason not to pick one of those three.
Bishop: https://www.amazon.com/Pattern-Recognition-Learning-Informat...
and
ESL: https://web.stanford.edu/~hastie/ElemStatLearn/
However what irks me a little, as it did with my university's introductory machine learning course, is the direct jump to the algorithms while omitting explaining the basis for the underlying cause why it is the right thing to do. Perhaps it's presumed that the reader knows the basic statistics behind the idea that we can make assumptions of the data and predict values due to central limit theorem. Machine learning itself then can be seen as a derivate of that as more free-form, heuristic approach to cases where the central limit theorem can't be applied or is too restrictive.
Am I speaking complete nonsense here? Please those wiser than me, tell me am I right?
From a theoretical standpoint, Machine Learning basically just explores what happens if you remove that assumption, and try to draw conclusions that apply no matter what the underlying distribution was.
The book does build up that theory. You need to understand the Probably Approximately Correct framework (e.g., how many data points do you need to be 90% sure your algorithm will generalize with less than 5% generalization error) and VC dimension (basically a measure of how much different algorithms overfit.) It doesn't even introduce an algorithm until chapter 8!
VC dimension is sort of the equivalent of O(n) style complexity for machine learning. You can use this framework to estimate the effectiveness of different algorithms from a theoretical standpoint given a certain amount of data. For example, Linear Regression's VC dimension grows as O(n) with the number of features, while Decision Trees grow as O(2^n) with depth. This means that Trees are much more likely to overfit on small data, but more likely to be find complex structure in larger datasets.
As others have mentioned this is a fairly theoretical take on machine learning which may not be useful if you just want to use a deep learning library. That said, I think there is a lot of value in having a deeper theoretical grasp of a topic even when practicing.
While feature engineering hasn't been a rigorously studied within the academic literature, this book does have a section on feature generation, which gives some practical tips once your data is in feature matrix form.
For those interested in even more, I work on a python library for automated feature engineering called, Featuretools (https://github.com/featuretools/featuretools/). It can help when your raw data is still too granular for modeling or comprised of multiple tables. We have several demos you can run yourself to apply it to real datasets here: https://www.featuretools.com/demos.