#56 Machine Learning vs Deep Learning

One of the most common and valuable IoT applications is predicting the failure of critical machinery.

Doing this prediction as a human might go like this: We identify some factors that could help us predict failure. We collect data to test our assumptions. Finally we build a model that does the final prediction using for example linear regression. It may even retrain itself as more data becomes available.

This isn’t really what we think of when we say ‘AI’ or ‘machine learning’ today, but actually it’s still machine learning by most definitions. What we tend to think of today is deep learning, where there is no longer any human process in identifying factors, but the AI can make its predictions using all data in any way it chooses.

The catch here is that deep learning requires training on enormous datasets. Previously we essentially used our human ability of deductive reasoning to narrow the search for a model. The appeal of deep learning lies in its ability to make predictions beyond the constraints of our imagination, yet this same quality necessitates the need for significantly larger amounts of data.

A good approach might be to start simple, collect data, grow the install base and consider deep learning as a second phase.