4 ms·
I signed up to HN just to reply. I suggest doing the MNIST dataset. The features are really simple (greyscale pixel values) yet it works very well. You can also
by mikehz 11y ago
I signed up to HN just to reply. I suggest doing the MNIST dataset. The features are really simple (greyscale pixel values) yet it works very well. You can also play with different classifiers to see how they behave (training time and classification performance).
This will let you setup a whole pipeline from feature selection (in this case just normalise, you can try 0 to 1 or -1 to 1 or subtract mean then divide by stddev, or don't normalize and see what happens) to training the model and evaluating its performance with cross validation. Then you can check your CV results by submitting to the leaderboard.
I took Andrew Ng's ML course then played with the MNIST dataset. I learnt heaps by doing this. Then I got carried away competing in real competitions. :) That's where more advanced feature selection came into play as well as making sure your CV split is representative of the test split.
I was using scikit-learn and just swapping classifiers in and out trying different ones as well as trying different parameters. You can even roll your own logistic regression if you want and see how regularisation affects performance etc.
- mrborgen 11y agoAwesome, thanks, I'll definitely test out that dataset! Btw, have you proceeded with any MOOC's after Andrew Ng's ML course? Anyone you'd like to recommend?
- bra-ket 11y agohttps://www.coursera.org/course/neuralnets https://www.coursera.org/course/neuralnets http://deeplearning.stanford.edu/tutorial/ http://deeplearning.stanford.edu/tutorial/ http://vision.stanford.edu/teaching/cs231n/syllabus.html http://vision.stanford.edu/teaching/cs231n/syllabus.html
- mikehz 11y agoSorry I didn't take any more courses after that. However I did continue to read every ML link that appeared on HN front page. Sometimes I'll go read the original paper to learn specific things such as DropConnect for NN (I saw slight improvement on MNIST with DropConnect but not on my other datasets). Same thing for domain knowledge for feature selection, just read relevant papers. Often I have to cross reference certain bits with other papers and reread 10 times before I understand the maths. It's still my goal to understand convolutional neural nets (and other topics like Bayesian statistics) but I've since stopped on ML and have been learning audio DSP instead. :) Too many things to learn...