I’ve been doing a bit of reading on machine learning lately as the field shows great potential for making a range of hard things easier.

One thing I have noticed pretty consistently in the books I’ve been going through is that they address the recognizer and statistical side of things in great detail but pass by the feature extraction side rather quickly.

It seems as if this approach misses some of the most challenging aspects involved in making a useful machine learning system. Given a blob of raw data, the identification and extraction of features that are suitable for processing by the ML system on the back end is non-trivial. Most of the sorts of data that I’d find interesting to process fall into this category.

Perhaps I’m missing something here, but it does not appear to me that feeding the entire photograph or audio stream to the machine learning algorithm is the intended approach. I’ll keep reading and sandboxing things (no real sandbox activity on this front yet as other priorities are ahead of ML in my queue). Hoping that I’ve missed something and this is less challenging that it appears from my current perspective.