Determining the source of neural-network dominance

Determine whether the current dominance of neural networks as a hypothesis language results from intrinsic properties of neural networks or from the substantially greater research effort, data, specialized hardware, and financial resources devoted to them.

Background

The paper notes that deep neural networks currently outperform other machine-learning approaches in scale and in their ability to learn from heterogeneous scientific data. Their success is associated both with intrinsic computational properties, such as differentiability and modular architectures, and with unusually large investments in data, hardware, and research.

The authors explicitly leave unresolved whether neural networks are inherently superior as a language for scientific hypotheses or whether their apparent dominance primarily reflects unequal development resources. Resolving this issue would clarify the prospects for alternative symbolic, probabilistic, or hybrid representations.