A Deep Learning Odyssey: An Invitation for Actuaries to Join this Journey.

While Deep Learning has shown remarkable success in the area of unstructured data like image classification, text analysis and speech recognition, there is very little literature on Deep Learning performed on structured/relational data. This investigation also focuses on applying Deep Learning on structured data because actuaries are more comfortable with structured data then unstructured data and so it can be a useful imitative step for other actuaries to apply Deep Learning to solve their own business problems. After extensive investigations, it does seem that Deep Learning has the potential to do well in the area of structured data. More specifically, we review Deep Learning, which is the most promising candidate so far for anomaly detection and for classification. We investigate class imbalance as it is a challenging problem for anomaly detection. In this report, Deep Multilayer Perceptron (MLP) was implemented using Theano in Python and experiments were conducted to explore the effectiveness of hyper-parameters. It was seen that increasing the depth of the neural network helped in detecting minority classes. Cost-sensitive learning technique was also observed to work quite well to deal with the class imbalance. One of the objectives of this project was to see the performance of dropout to regularize and avoid over-fitting in the network. Our conclusion is that while dropout with techniques worked for unstructured data, adding dropout to structured data does not seem to give any improvement. We also situate Deep Learning in its proper strategic context; highlight its applications and limitations as well as on-going developments in Deep Learning.