Deep learning has become the dominant modeling paradigm in machine learning. It has been spectacularly successful in application areas ranging from speech recognition, image analysis, natural language processing, and information retrieval. But a number of important challenges remain un(der)solved, such as data efficient deep learning, energy efficient deep learning and visualizing deep neural networks. In this talk I will address the problem of “data efficient deep learning” through three distinct approaches:
1) Combining generative probabilistic (graphical models) with deep learning using variational auto-encoders (w/ D. Kingma),
2) Bayesian deep learning using variational approximations based on matrix-normal distributions on random matrices (w/ C. Louizos)
3) Exploiting symmetries using Group-equivariant CNNs (w/ T. Cohen)