Deep learning has provided means to model complicated systems using very large amounts of data. I work in collaboration with hospitals to gather image and sensor data, with annotation to build models for classification tasks. The challenge is, after building the models, to find out the essential features that predict the classification. If successful, one can get hold on some part of the “dark knowledge” from decade long experience of professionals, which they may find hard to quantify and explain.
Also, when data is sparse, simulations of physical systems can be used to produce datasets for training regression models and classifiers. One example is utilizing convolutional neural networks to estimate material parameters in ultrasound tomography.
Finally, when the models are trained and perform well enough, the challenge is to be able to simplify these in a way that makes them run even in embedded systems without significant loss of accuracy.