Study quantifies how deep learning model reuse is reshaping scientific research

Researchers have conducted the first large-scale empirical investigation into how natural scientists are adopting pre-trained deep learning models (PTMs) to reduce the computational burden and expense of developing AI systems from scratch.

Why model reuse matters#

Training deep learning models demands significant computational resources and technical expertise—barriers that have historically limited their adoption in academic science. To address this, researchers in natural sciences have increasingly turned to reusing pre-trained models, a practice endorsed by the software engineering community as a cost-effective strategy.

What the data shows#

Analyzing 17,718 peer-reviewed, open-access papers, the researchers found:

• Biochemistry, Genetics and Molecular Biology has become the leading adopter of PTM reuse among natural science disciplines
• Adaptation emerged as the dominant reuse pattern, where scientists fine-tune existing models for their specific applications
• The testing phase of the scientific process has experienced the most significant impact from PTM integration

Implications for science#

These findings suggest that deep learning adoption patterns in academic research are following established software engineering best practices. Rather than training models independently, researchers are now building on shared foundations, potentially accelerating discovery while making advanced computational methods accessible to scientists without massive AI infrastructure budgets.

The work represents the first systematic quantification of how pre-trained models have fundamentally reshaped the scientific process across multiple disciplines.

Source: arXiv cs.SE