Deep Learning Starter

Artificial Intelligence / Data Science / Machine Learning

Deep learning has garnered a lot of attention recently, even from Google, who make their largest European acquisition when they bought DeepMind for $400 million at the beginning of the year. Part of this attention can be explained by the general rise in popularity of predictive analytics, but the recognition of the improved performance of deep learning, as compared to normal machine learning, is certainly adding fuel to the fire.

John Platt (at Microsoft Research) has given a good short explanation of the technique. In order to recognise patterns using conventional machine learning, the features of the patterns have to be picked out manually and assigned a weighted importance. Identifying the features in the raw data is actually very hard and a benefit of deep learning is that it takes away this manual step. In deep learning, the most useful features are discovered during the training process.

It is a very new technique and research area. There are only about 50 deep learning experts worldwide, and many are still graduate students! Here is a great free book by Michael Nielsen on deep learning with nice examples to work through. And a Google Tech Talk on deep learning by Yoshua Benigo.

Despite it being a very new field, it has already started generating impressive results. Christopher Olah has reviewed deep learning applied to natural language processing. Facebook created DeepFace, a deep learning feature with 97.25% accuracy in detecting whether two faces in unfamiliar photos are of the same person. The deep learning part of DeepFace consists of nine layers of simple simulated neurons, with more than 120 million connections between them. It is pretty awesome.

Even though deep learning seems to be such as a powerful technique, data science practitioners tend to be careful when using it. Despite the improved results, the algorithms are often uninterpretable, so it is incredibly difficult to know if anything is going wrong or how to fix the model if something has gone wrong. Nonetheless, this is a very exciting space and worth watching carefully.

Share on: Twitter, Facebook, LinkedIn or Google+