Many colleagues and graduate applicants choose to work on deep learning or an application of it to a problem in vision, imaging, robotics, bioinformatics, graphics or another domain. Our community is excited about applying models (typically pre-trained on ImageNet, Places, or related databases) to datasets, using existing software toolkits and GPU enhanced machines. Then, there is a bit of tweaking of the underlying architecture, hyperparameter tuning, and often some form of data augmentation. A button is then pressed, results are recorded on a benchmark, and a new paper is written. Dozens of papers a year in venues including cvpr, eccv, iccv, miccai and nips, follow this pipeline. Each year we come up with new heuristics to modify the architectures, with a system level diagram to articulate the advances we have made, replete with hacks that few care to analyze or really understand. Industry and the media are lapping this up, believing us to be at the dawn of a new age in artificial intelligence.
Why have we abandoned good old fashioned science?
In our computer vision course the lecture on Deep Learning includes a box with an image on the left and the label "cat" on the right. That box is black.