Deep Neural Networks Revolution

Statisticians look at prediction as a wighted average over extended attribute space. Examples range from highly interpretable models like ARIMA, GARCH, linear regression, to more sort of black box techniques such as principal component regression, kernel method, smoothing splines, regression tree, random forest, and project pursuit. However, deep neural networks provides a totally different insight: repeated projection-cuts trained over a large set of data. Neural networks has been around for a while, but cheap parallel computing hardware such the GPUs and cloud servers has given this method a considerable push ahead, specially for big data. I argue that deep learning is not only an artificial intelligence tool, but can be considered as a strong candidate in all modeling problems that aims at promoting model prediction accuracy over model interpretation power. It is still difficult to answer why deep learning works so well, but I will provide some intuitions that emphasizes on projection-cut as a remedy to the curse of dimensionality. The performance of deep learning on certain applications is largely ahead of other methods, and this is why most industries are interested to invest on adopting this method for their prediction challenges. I see a great potential for deep learning in insurance industry to automate customer experience, fraud investigation, claim processing, and so on.