R_Saenet

Talk Outline: The SAENET package is designed to facilitate deep learning and pre-training of feed-forward neural networks in R. It builds upon the 'autoencoder' package written by Eugene Dubossarsky and Yury Tyshitsky, which conducts single-layered deep feature learning using backpropagation, to allow multi-layered deep learning of latent features in datasets of an arbitrary size.

Using this implementation, it is possible to easily create compressed representations of data using features arrived at through deep learning or alternatively to pre-train feed-forward neural networks using the well-established 'neuralnet' package available in R.

In this talk we will discuss the conceptual underpinnings of autoencoders, how they work in the single-layered and multi-layered cases and finally how they can be used to pre-train neural networks and the benefits of doing so. By the end of this, the general idea of deep learning and how it can be used to understand and analyse data will hopefully be a touch clearer to the audience.