A new textile production line has 40% of the sensor data a mature line has collected. Not enough to train a deep neural network from scratch. But the physics of textile manufacturing — the relationship between loom speed, thread tension, energy consumption, and output quality — is shared across lines. The knowledge exists. It's just in the wrong place.
Ensemble Deep Transfer Learning (arXiv:2601.12663) trains models on data-rich production lines first, then transfers the learned representations to data-poor lines. The improvement: 5.66% in prediction accuracy, 3.96% in robustness, specifically in the 20–40% data availability regime where conventional models fail.
The structural insight is that what transfers isn't the specific sensor readings but the relational structure between them. Thread tension relates to energy consumption the same way regardless of which loom it's measured on. The model learns the physics encoded in the correlations, not the absolute values. When applied to a new line, it needs only enough local data to calibrate the scale — not to rediscover the relationships.
The ensemble component matters because transfer from a single source model can overfit to the source domain's idiosyncrasies. Aggregating multiple adapted models averages out the source-specific noise, leaving only the generalizable structure. More donors, less bias.
This is the data-efficiency argument for transfer learning made concrete: in manufacturing, data is expensive because it means production time, sensor maintenance, and quality inspection. Borrowing knowledge from existing lines isn't a shortcut. It's the only economically viable path.