ABSTRACT
AI and especially deep learning are increasingly being adopted as tools into the scientific workflow, yet in some ways DL is antithetical to established scientific principles. One such principle is Occam’s razor, the preference for models that are as simple as possible - deep learning models on the other hand are enormously over-parameterized, in stark opposition to this principle. Deep learning has managed to avoid overfitting, using approaches other than parsimony of parameters, thereby addressing at least one practical issue that is naturally avoided by Occam’s razor. Several research efforts to make neural networks more explainable and interpretable are underway, however for the moment they still remain more of a ‘black box’ than regular scientific models. Yet, the quality and breadth of deep learning results are hard to dismiss, and this state of affairs is compelling scientists to re-examine basic tenets and their trade-offs. Among other things, end-to-end deep learning offers a noteworthy value proposition in addressing the ‘data ingest problem’, the difficulty that is encountered in translating observation data into a form that a classical model can utilize, as well as the equivalent problem on the output side of the modeling process.
