Network Transplanting

This paper focuses on a novel problem, i.e., transplanting a category-and-task-specific neural network to a generic, distributed network without strong supervision. Like playing LEGO blocks, incrementally constructing a generic network by asynchronously merging specific neural networks is a crucial bottleneck for deep learning. Suppose that the pre-trained specific network contains a module $f$ to extract features of the target category, and the generic network has a module $g$ for a target task, which is trained using other categories except for the target category. Instead of using numerous training samples to teach the generic network a new category, we aim to learn a small adapter module to connect $f$ and $g$ to accomplish the task on a target category in a weakly-supervised manner. The core challenge is to efficiently learn feature projections between the two connected modules. We propose a new distillation algorithm, which exhibited superior performance. Our method without training samples even significantly outperformed the baseline with 100 training samples.