How statistics can help us do “data science”: examples in deep learning and graph analysis

“Data science”, beyond being a buzzword, has indeed become a major focus in a variety of scientific communities. Even though, it has not yet been properly defined, it is a consensus that “data science” involves an intersection of statistics, computer science and domain knowledge. In this talk, we discuss two examples where statistics plays a major role in performing “data-scientific” enquires.

In the context of analysing neural networks, we will show how tools from statistics can be used to answer questions such as “why training neural networks gets harder as the depth of the network increases?” and “what could potentially be appropriate strategies for initialization of weights in the training process?”.

We then discuss an example where statistics can help study large-scale graph spectral clustering problems. In particular, we show that, under certain statistical assumptions, given an existing clustering of a large graph, one can efficiently and yet accurately, perform an out-of-sample extension to observations not seen in the initial clustering procedure.