Machine learning is like the game of go.

Machine learning is like the game of go.

- David Rolland

The most popular machine learning algorithms are based on statistical theories. In the recent years we have seen the coming of big data which is simply the only improvement potential to these algorithms.

I don’t belong to this school of thought. I believe that applying logic will give you better results. More data is always welcome to improve the accuracy of an algorithm but more data doesn’t necessarily have to be as much as “Big Data”. By using logic and better selected data and features you can go farther. This where I see machine learning looks like the game of go. In the game of go you need to place your stone at the right place to gain the most territory. Put your stones too close from one another and your enemy will take over most of the territory, play them too far away and your opponent will invade the territory in-between. In machine learning you need to use a sample of data carefully chosen to cover most of what the real world covers. Your sample must be chosen just like you choose where to put your stones in the game of go. This way, you don’t need millions, or even billions, of samples to get an equivalent accuracy and getting this smaller amount of samples uses way less resources.

The same can be said about machine learning features. They need to be relevant to the machine learning problem. Instead of using thousand of low quality features, you better use less features but more carefully selected ones. These better feature might be longer to compute but if you’re using logic instead of statistics you will not have to compute every feature for each tested data point as your model is able to return it’s classification decision without depending on all features.