Ask HN: How to get started with machine learning?
How should a software engineer with no machine learning background get started on the subject?
Do you think that getting started by learning a framework like TensorFlow is a good idea or should I gain a background knowledge first?
130 comments
[ 5.0 ms ] story [ 139 ms ] threadIf you like to study/read: the famous Coursera Andrew Ng machine learning course: https://www.coursera.org/learn/machine-learning
If you just want course materials from UC Berkeley, here's their 101 course: https://news.ycombinator.com/item?id=11897766
If you want a web based intro to a "simpler" machine learning approach, "decision trees": https://news.ycombinator.com/item?id=12609822
Here's a list of top "deep learning" projects on Github and great HN commentary on some tips on getting started: https://news.ycombinator.com/item?id=12266623
If you just want a high level overview: https://medium.com/@ageitgey/machine-learning-is-fun-80ea3ec...
Can you know nothing about ml, ai, data analysis, and stats then give tensor flow some input and it will give you some input and pretty much apply it to your app?
Or do you have to know these subjects before even starting tensor flow?
It's OK to jump in and try it without having background information. See how far you get and start researching when you hit a wall or find sudden interest.
Thanks for the link
http://academictorrents.com/browse.php?search=machine+learni...
Unlike some of the other complicated tools, sklearn is just a "pip install" away and includes all sorts of examples of different problems. Classification? Regression? Clustering? Representation learning? Perceptual embedding? Odds are, some part of sklearn covers all of that.
This means that you can set up a train and and test set and swap in and out random forest, svm, naive bases, logistic regression, and various others.
Read about them one by one, try to understand the algorithms generally, test them out, see how they perform differently on different data sets.
It all depends on how you like to approach a new subject, but I think this is more fun and motivating than going straight into the mathematics behind the algorithms right away (which is more along the lines Andrew Ng's excellent course). I'd say once you're into it and using the algorithms, then dig deeper into the core mathematics, you'll have a better context for it.
[1] – https://en.wikipedia.org/wiki/Iris_flower_data_set
The process of learning could be defined as a task of extraction of relevant information (knowledge) about reality (shared environment) not mere accumulation of a fancy nonsense or false beliefs.
And reality like: The actual sinking of the Titanic?
If your model concludes that nobility, traveling first class, close to the exits, without family, has a higher chance of surviving, then this is fancy nonsense or a false belief?
You make a really strange case for your view.
In real life it is also very rare to have free pickings of the variables you want. Some variables have to substituted with available ones.
The Titanic story is to make things interesting for beginners. One could leave out all the semantics of this challenge, anonymize the variables and the target, and still use this dataset to learn about going from a table with variables to a target. In fact, doing so teaches you to leave your human bias at the door. Domain experts get beaten on Kaggle, because they think they need other variables, or that some variables (and their interactions) can't possibly work.
Let the data and evaluation metric do the talking.
It shows only that given set of variables (observable and inferred) could be used to build a model. The given data set is not descriptive, because it does not contain more relevant hidden variables, so any predictions or inferences based on this data set are nothing but a story, a myth made from statistics and data.
[1] https://en.wikipedia.org/wiki/Testing_hypotheses_suggested_b...
And if you see classification as a form of hypothesis testing, then cross-validation is a valid way of testing if hypothesis holds on unseen data.
This would be like sampling all coins from my pockets and thinking you could build a predictive model of year printed to value of coin. Probably could for the change I carry. Not a wise predictor, though.
This dataset is more in line with what you are looking for: https://www.kaggle.com/saurograndi/airplane-crashes-since-19...
That sounds a bit iffy. A domain expert should really know what they're talking about, or they're not a domain expert. If the real deal gets beaten on Kaggle it must mean that Kaggle is wrong, not the domain expert.
Not that domain experts are infallible, but if it's a systematic occurrence then the problem is with the data used on Kaggle, not with the knowledge of the experts.
I mean, the whole point of scientific training and research is to have domain experts who know their shit, know what I mean?
Notably: https://www.kaggle.com/c/MerckActivity
> Since our goal was to demonstrate the power of our models, we did no feature engineering and only minimal preprocessing. The only preprocessing we did was occasionally, for some models, to log-transform each individual input feature/covariate. Whenever possible, we prefer to learn features rather than engineer them. This preference probably gives us a disadvantage relative to other Kaggle competitors who have more practice doing effective feature engineering. In this case, however, it worked out well.
http://blog.kaggle.com/2012/11/01/deep-learning-how-i-did-it...
> Q: Do you have any prior experience or domain knowledge that helped you succeed in this competition? A: In fact, no. It was a very good opportunity to learn about image processing.
http://blog.kaggle.com/2016/09/15/draper-satellite-image-chr...
> Do you have any prior experience or domain knowledge that helped you succeed in this competition? I didn't have any knowledge about this domain. The topic is quite new and I couldn't find any papers related to this problem, most probably because there are not public datasets.
http://blog.kaggle.com/2015/09/16/icdm-winners-interview-3rd...
> Do you have any prior experience or domain knowledge that helped you succeed in this competition? M: I have worked in companies that sold items that looked like tubes, but nothing really relevant for the competition. J: Well, I have a basic understanding of what a tube is. L: Not a clue. G: No.
http://blog.kaggle.com/2015/09/22/caterpillar-winners-interv...
> We had no domain knowledge, so we could only go on the information provided by the organizers (well honestly that and Wikipedia). It turned out to be enough though. Robert says it cannot happen again, so we’re currently in the process of hiring a marine biologist ;).
http://blog.kaggle.com/2016/01/29/noaa-right-whale-recogniti...
> Through Kaggle and my current job as a research scientist I’ve learnt lots of interesting things about various application domains, but simultaneously I’ve regularly been surprised by how domain expertise often takes a backseat. If enough data is available, it seems that you actually need to know very little about a problem domain to build effective models, nowadays. Of course it still helps to exploit any prior knowledge about the data that you may have (I’ve done some work on taking advantage of rotational symmetry in convnets myself), but it’s not as crucial to getting decent results as it once was.
http://blog.kaggle.com/2016/08/29/from-kaggle-to-google-deep...
> Oh yes. Every time a new competition comes out, the experts say: "...
https://en.wikipedia.org/wiki/Ant_colony_optimization_algori...
https://en.wikipedia.org/wiki/Artificial_ants
Most of the models mimic and simulate (very naively) that observable behavior, not its origin.
When people cite "the map is not the territory" they mean this. Simulation is not even an experiment. It is mere an animation of a model - a cartoon.
Simulation can be a very beneficial experiment. See for instance: https://papers.nips.cc/paper/5351-searching-for-higgs-boson-...
Simulations are not experiments. It is an animation of a formalized imagination, if you wish.
If you work through the data, you'll find things like women, children and first class passengers had a higher survival rate than men with lower class tickets[1].
This matches exactly the stories of what happened: Staff took first class passengers to the lifeboats first, then women and children. Then they ran out of lifeboats.
So the data shows correlation, and eye-witness accounts shows causation. That's close to the ideal combination: eyewitness accounts can be unreliable because we can't know how widespread they are, and correlation doesn't show causation.
But the combination of them both is pretty much the best case for studying something which can't be replicated.
[1] See examples like https://www.kaggle.com/omarelgabry/titanic/a-journey-through...
But any attempt to frame it as a "prediction", an accurate model of the event or adequate description of reality is just nonsense.
To call things by its proper names (precise use of the language) is the foundation of the scientific method. This is mere oversimplified, non-descriptive toy model of one aspect of historical event, made from of statistics of partially observable environment. A few inferred correlations reflects that there was not a total chaos, but some systematic activity. No doubt about it. But this is absolutely unscientific to say anything else about the toy model, let alone claim that any predictions based on it have any connection to reality.
There is clear correlation between gender and survival rates. Given the data, a decent prior would absolutely take that into account.
Yes, there are other factors. But the foundation of statistical models is simplification, and descriptive statistics are an important foundation of that.
In any case, it isn't exactly clear that there are magical hidden factors which predicted survival. It appears you maybe unfamiliar with the event, because basically those who got into a lifeboat survived, and those who didn't, didn't survive.
To quote Wikipedia:
Almost all those who jumped or fell into the water drowned within minutes due to the effects of hypothermia.... The disaster caused widespread outrage over the lack of lifeboats, lax regulations, and the unequal treatment of the three passenger classes during the evacuation..... The thoroughness of the muster was heavily dependent on the class of the passengers; the first-class stewards were in charge of only a few cabins, while those responsible for the second- and third-class passengers had to manage large numbers of people. The first-class stewards provided hands-on assistance, helping their charges to get dressed and bringing them out onto the deck. With far more people to deal with, the second- and third-class stewards mostly confined their efforts to throwing open doors and telling passengers to put on lifebelts and come up top. In third class, passengers were largely left to their own devices after being informed of the need to come on deck.
Even more tellingly:
The two officers interpreted the "women and children" evacuation order differently; Murdoch took it to mean women and children first, while Lightoller took it to mean women and children only. Lightoller lowered lifeboats with empty seats if there were no women and children waiting to board, while Murdoch allowed a limited number of men to board if all the nearby women and children had embarked
All this behavior matches exactly what the model tells us about the event.
I'd be very interested if you can point to something specific that is wrong about it.
All models are wrong, but some are useful.
Exactly.
All predictions are wrong and make no sense for partially observable, multiple causation, mostly stochastic phenomena. It will never be the same.
The Titanic's sister ship (the Brittanic) was torpedoed during WW1 and sunk. However, the lesson of the Titanic (too few lifeboats) had been learnt, and only 26 people died.
I don't know what point you are trying to make - yes, I agree that history never repeats, but lessons can be learnt from it, and they can be quantified and they can be useful.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1291331/
OK, tell me, please, what it is that you can predict? That some John Doe, having the first class ticket in a cabin next to the exit would survive the collision of the next Titanic with a new iceberg? That being a woman gives you better chances to secure a seat in a lifeboat? What is the meaning of the word "predict" here?
This happened because they made a _statistical_ model of the Titanic disaster, and learned from it? Like, they actually crunched the numbers and plotted a few curves etc, and then said "aha, we need more boats"?
I kind of doubt it, and if it wasn't the case then you can't very well talk about a "model", in this context. It's more like they had a theory of what factor most heavily affected survival and acted on it. But I'd be really surprised to find statistics played any role in this.
No - statistics as the discipline that we think of today wasn't really around until the work of Gosset[1] and Fisher[2] which was done a few years after this.
I'm sure you noted that I was very careful with what I claimed: "the lesson of the Titanic (too few lifeboats) had been learnt".
These days we'd quantify the lesson with statistics. Then, they didn't have that tool.
Instead, we have testimony[3] relaying the same story: Just one question. Have you any notion as to which class the majority of passengers in your boat belonged? - (A.) I think they belonged mostly to the third or second. I could not recognise them when I saw them in the first class, and I should have known them if there were any prominent people. (Q.) Most of them were in the boat when you came along? - (A.) No. (Q.) You put them in? - (A.) No. Mr. Ismay tried to walk round and get a lot of women to come to our boat. He took them across to the starboard side then - our boat was standing - I stood by my boat a good ten minutes or a quarter of an hour. (Q.) At that time did the women display a disinclination to enter the boat? - (A.) Yes."
So yes, I agree - it was a theory, which our modern modelling tools can show matched well with what the statistics showed happened.
My whole point is that this is very useful, unlike the OP who dismissed it as useless.
[1] https://en.wikipedia.org/wiki/William_Sealy_Gosset
[2] https://en.wikipedia.org/wiki/Ronald_Fisher
[3] http://www.titanicinquiry.org/BOTInq/BOTInq33Header.php
Now, it so happens that that correlated heavily with class. But, not as much as with sex. Though, there were some places where being male hurt your chances (as you point out in the one officer not allowing men on boats), by and large these were secondary and correlated with success, not predictors of it.
The first lesson is a classifier to separate apples and oranges.
https://www.youtube.com/watch?v=cKxRvEZd3Mw&list=PLOU2XLYxms...
I read all of these architectures in research papers, but I'd really love to learn how to start iterating on them for a particular domain.
If you don't understand something in the book, back up and learn the pre-reqs as needed.
http://web.stanford.edu/~hastie/ElemStatLearn/printings/ESLI...
http://www-bcf.usc.edu/~gareth/ISL/
Personally, I studied Duda & Hart's pattern recognition [1] and Casella & Berger's statistics text [2] simultaneously. This took about the equivalent of 2 semesters. Duda's text gets the main ideas across without being as heavy on the probability theory / stats.
Afterwards, I studied "Elements ..." by Hastie et al., which was far more readable after going through Casella & Berger's text. Now Hastie et al. is my go-to reference. I also should note that this all assumes that you also have the requisite math background: up to calc 3, linear algebra, and maybe some exposure to numerical methods (in particular, optimization).
[1]: https://books.google.com/books?id=Br33IRC3PkQC&lpg=PP1&pg=PR...
[2]: https://books.google.com/books/about/Statistical_Inference.h...
Everyone keeps linking ESL, but really ISLr is much easier to understand, provides more important clarifying context, and covers more or less the same information.
ESL is more like a reference and prototype for ILSr
http://www-bcf.usc.edu/~gareth/ISL/
Also Fermat's Library is going to be annotating the book, which should make it even more accessible: http://fermatslibrary.com/list/neural-networks-and-deep-lear...
and then continue with https://www.coursera.org/learn/neural-networks/
I strongly advice for:
- using Python in the interactive environment Jupyter Notebook,
- starting with classical machine learning (scikit-learn), NOT from deep learning; first learn logistic regression (a prerequisite for any neural network), kNN, PCA, Random Forest, t-SNE; concepts like log-loss and (cross-)validation,
- playing with real data,
- it is cool to add neural networks afterwards (here bare TensorFlow is a good choice, but I would suggest Keras).
Links:
- http://www.r2d3.us/visual-intro-to-machine-learning-part-1/
- http://hangtwenty.github.io/dive-into-machine-learning/
- https://github.com/leriomaggio/deep-learning-keras-euroscipy...
http://cs231n.github.io/
Really great content from Andrej and his coworkers. This guy is great.
You can easily find all classes videos on YouTube too.
1. Machine Learning by Andrew Ng (https://www.coursera.org/learn/machine-learning) /// Class notes: (http://holehouse.org/mlclass/index.html)
2. Yaser Abu-Mostafa’s Machine Learning course which focuses much more on theory than the Coursera class but it is still relevant for beginners.(https://work.caltech.edu/telecourse.html)
3. Neural Networks and Deep Learning (Recommended by Google Brain Team) (http://neuralnetworksanddeeplearning.com/)
4. Probabilistic Graphical Models (https://www.coursera.org/learn/probabilistic-graphical-model...)
4. Computational Neuroscience (https://www.coursera.org/learn/computational-neuroscience)
5. Statistical Machine Learning (http://www.stat.cmu.edu/~larry/=sml/)
If you want to learn AI: https://medium.com/open-intelligence/recommended-resources-f...
So you could start with some really simple example code for demand forecasting but where you put in your data and your signals. In this way you can learn what you need to solve a particular problem, 'getting lucky' from only having to adapt examples. Sure it might be nice to learn all the fundamentals first but it is sometimes nice to scratch an itch, every company has plenty, choose one and see how far you get and learn along the way.
https://metacademy.org
"Your package manager for knowledge".
(mostly focused around ML)
Forget about the code part. It's the least difficult part.
With modern tools and frameworks you can start learning and applying what you know on practice almost immediately.
Check out Keras and the book "Deep Learning with Python"[1]. They have enabled me to train my first ANN in 2 days, and get to the point of building a MNIST recognizer in a month(and I was reading it pretty slowly).
Sure, if you're coding it from scratch and must understand every signle detail, you do need like 10 years and 3 PhD's. But that's not a wise way to learn.
I recommend to take the simplest tools, and apply them to practical projects immediately. That will give you the general overview of how things work, and then you will learn the details as needed.
[1] https://machinelearningmastery.com/deep-learning-with-python...
Yes, you can take a library and implement it in 10 minutes, but then you're really not learning machine learning, are you?
I will argue you do not need four years of math by any stretch, though. The stumbling block will be notation more than anything else. Relatively basic calculus and linear algebra will suffice.
They were right about one thing: the code is the least important part.
As an undergrad, I was doing all those easy ML tutorials and took an undergrad level ML course. I thought I would be useful in actual practice, but knowing the whats/hows of neural nets/clustering/etc. is not enough. Feature engineering/math is the most difficult part. In a corporate setting, if it was a straight forward solution, you wouldn't be doing that work because the solution would be trivial and already implemented.
As an engineer with only a bachelors on a ML team full of PhDs there is a definite difference in skill. I've been reduced to a monkey (a content one) that works on the data pipeline. Learning to deal with real world ML problems would take me years of work that I am not sure I would be willing to do, especially when the pay increases per effort expended learning ML is much lower than with regular software/distributed systems/etc..
On the interest part, you're right that I would never have tried to learn ML if I had known the amount of work that is required to actually be good or if I tried learning the math first. That's the real world though. The useful ML engineers did learn the math. The efficient way to learn ML is to learn the math/statistics first.
For special applications, it is totally OK to learn as you go.
http://blog.digitalmind.io/post/artificial-intelligence-reso...
I too felt like ML is something new to try, but the lack of real world use cases on a small scale ( not google, Microsoft, ... ) Has kept me from trying/doing.
I only saw the farm with image recognition for vegetables as an example for now.
Anyone has other examples?
Numerous ML competitions also provide enough fun to get started.
Mathematical Monk - https://www.youtube.com/user/mathematicalmonk#p/c/0/ydlkjtov... (includes a probability primer)
Awesome Courses - https://github.com/prakhar1989/awesome-courses - its a very extensive list of university courses including subjects apart from Machine Learning as well
Programming Collective Intelligence - http://www.amazon.com/programming-collective-intelligence-bu... - heard very good reviews about this
Many other resources available apart from the above. You can access more such resources at http://www.tutorack.com/search?subject=machine%20learning
I think its a good idea to go through one or more beginner level courses like that offered by Andrew Ng on Coursera and then do an actual project.
[Disclaimer - I work at tutorack.com mentioned in the comment]
Instead, learn decision trees and more importantly enough statistics so you aren't dangerous.
Do you know what the central limit theorem is and why it is important? Can you do 5-fold cross validation on a random forest model in your choice of tool?
Fine, now you are ready to do deep learning stuff.
The reason I say not to do neural networks first is because they aren't very effective with small amounts of data. When you are starting out you want to be able to iterate quickly and learn, not wait for hours for a NN to train and then be unsure why it isn't working.
"All of statistics" is really a great book if you have time work through he exercise.
http://www.inference.phy.cam.ac.uk/itila/book.html (freely accessible online)
https://www.openintro.org/stat/textbook.php
Of course it's important to get a broad horizon eventually but starting with the theory without the applications is not how most humans learn best. Learning by doing is.
The problem with diving into neural networks is that they are slow to train (with large amounts of data anyway), and difficult to debug. This means it isn't really a great place to start.
There are basic things I think you must know before jumping into a framework or int any specific algorithm. First thing you probably will have to do is to collect the data and clean it. In order to do this correctly you need some basic statistics. For example you need to know what is a gaussian distribution and collect samples in a way that are representative of your problem. Then you may need to clean the samples, remove outlines, complete blank data, etc. So it is basic you know some statistics to do this right.I have seem people with a lot of knowledge of tools than then they are not able to create a train/test/validation set correctly and the experiment is completely invalid from here no matter what you do next (http://stats.stackexchange.com/questions/152907/how-do-you-u..., https://www.youtube.com/watch?v=S06JpVoNaA0&feature=youtu.be ). You also need to know how are you going to test your results, so again you need to know how to use a statistical test (f-test, t-test). So first thing, jump into statistics to understand your data.
The next step I think is to know some common things in machine learning as the no free lunch theorem, curse of dimensionality, overfitting, feature selection, how to select the current metric to asses your model and common pitfalls. I think the only way to learn this is reading a lot about machine learning and making mistakes by your own. At least now you have some things to search in google an start learning.
The third step would be to understand some basic algorithms and get the feeling of the type of algorithms, so you know when a clustering algorithm is needed or your problem is related classification or with prediction. Sometimes a simple random forest algorithm or logistic regression is enough for your problem and you don't need to use tensorflow at all.
Once you know the landscape of the algorithms I think it is time to improve your maths skills and try to understand better how the algorithms works internally. You might not need to know how a deep network works completely, but you should understand how a neural network works and how backpropagation works. The same with algorithms as k-means, ID3, A*, montecarlo tree search or most popular algorithms that you are probably are going to use in day to day work. In any case you are going to need to learn some calculus and algebra. Vectors, matrix and differential equations are almost everywhere.
You would probably have seen some examples when learning all the stuff I talked about, then it is time to go to real examples. Go to kaggle and read some tutorials, read articles about how the community of kaggle has faced and winning the competitions. From here is just practice and read.
You can jump directly into a framework, learn to use it, have 99% accuracy in your test and 0% accuracy with real data. This is the most probably scenario if you skip the basic things of machine learning. I have seen people doing this and end up very frustrated because they don't understand how their awesome model with 99% accuracy doesn't work in the real world. I have also seen people using very complex things as tensorflow with problems that can be solved with linear regression. Machine learning is a very broad area and you need maths and statistics for sure. Learn a framework is useless if you don't understand how to use it and it might lead you to frustration.
[1] http://www.cs.waikato.ac.nz/ml/weka/
Now practical: I think the best way to learn is pick an algorithm & representation and implement it in your favorite language. Bonus if you have your own language to work with.
I would start looking into Decision Trees first, implement them and then implement some use cases(, which follow after implementing them). Do this for other approaches, like ANN, which you can have it beat you at checkers which is strangely satisfying.
But keep in mind Minsky. I think he is like Archimedes doing "Calculus"-type approaches without fully realizing. Maybe you could be Newton?