Andrew Ng's course notes are 90% of the way to being a textbook, and are an incredible resource, especially if you're going to watch the lectures.<p><a href="http://www.stanford.edu/class/cs229/materials.html" rel="nofollow">http://www.stanford.edu/class/cs229/materials.html</a>
God, whatever you do please don't start by picking a few datasets at UCI ML site and using R packages to play with the data.<p>An ideal approach will be to:<p>- Pick any programming language and start off with plain regression. It may look simple but this will become fantastic base going forward<p>- Generate a synthetic data set and apply your freshly written regression on it<p>- Expand your toolkit to include test and training data set generation and calculation of ROC curves and confusion tables<p>- Add logistic regression, regularizers and other advanced regression models to the toolkit<p>- Use a real world dataset and develop multiple different models. And pick the best model (choosing the right model itself is a big task in itself)<p>- Then try coding Neural Networks, SVM, etc.