Machine Learning - Old Fish in New Paper

When NOT to use deep learning? Pablo Cordero’s post, Jeff Leek's Simply Stats Blog, and a rebuttal from Andrew Beam make good points comparing the two approaches, but I’m not so interested in the issues related to sample size; I need to understand what’s happening under the hood. 

I choose methods based on understanding and interpretability. I spend more time on defining the problem and determining the appropriate questions than on the actual computation. I need statistical models to give me a defendable understanding of the underlying processes; they provide confidence intervals, optimization opportunities, diagnostics and graphical methods. This overrides any overarching drive to get one last 0.5% increment on performance. Even if I have the best model in the world, the first question in the meeting is usually about how we should change the underlying process to increase, decrease or otherwise improve or optimize the process.

Over twenty years ago when neural nets were (once again) the latest thing, I found an article, Neural Networks and Statistical Models by Warren S. Sarle at the SAS Institute, where he showed the relationship between the two approaches. As he puts it, “translates neural network jargon into statistical jargon”. The paper is a great summary of various neural net models with their equivalent statistical models. For example, this figure illustrates how a multilayer perceptron (MLP) is equivalent to multivariate multiple nonlinear regression.

This gave me a foundation for investigating the limitations and capabilities of new techniques.

tl;dr:

Learn the detailed underlying logic and assumptions behind existing techniques before jumping on the latest bandwagon. You might be using an old reliable technique in disguise.