Deep Learning with Small Datasets: Tips, Tricks, and Cautionary Tales

Many modern applications of deep networks are generative or discriminatory tasks that leverage large datasets and computational power to achieve state-of-the-art results on tasks in important, wide-reaching regimes. However, some applications are more data-constrained. This talk will focus on uses of neural networks in medical imaging and remote sensing, two application areas where ground-truth is hard to come by, but the potential benefits of the deep learning revolution are enormous. I will discuss modern deep learning techniques for image reconstruction, how to deal with small datasets, explore recent work that provides insight into ways deep learning can fail in the field, and discuss promising directions for the future.