Guzdial, Riedl propose recombining neural nets without retraining

"Conceptual expansion," proposed by Matthew Guzdial and Mark O. Riedl, is a method for building new neural network models by recombining already-trained ones, with no additional backpropagation. The authors posted the paper to arXiv in February 2018 and revised it that September.

They describe conceptual expansion as "a general representation for reusing existing trained models to derive new models without backpropagation," and evaluate on few-shot versions of two tasks: image classification and image generation. On those tasks, the method can "outperform standard transfer learning approaches," the authors report.

The framing leans on human cognition. "One of the defining characteristics of human creativity is the ability to make conceptual leaps, creating something surprising from typical knowledge," the abstract reads. Deep networks, by contrast, "often struggle to handle cases outside of their training data, which is especially problematic for problems with limited training data."

The paper runs nine pages with a single figure. The abstract does not name the datasets used, the few-shot sample sizes, or the specific transfer-learning baselines the authors compared against.