Can you reuse part of one CNN to build or fix another? Our ICSE 2022 paper makes CNNs modular.

May 24, 2022

Building a convolutional neural network almost always starts from scratch, with training that costs time and energy. Yet much of what one network learns, such as the ability to recognize a particular class, overlaps with what other networks need.

In this paper (ICSE 2022), Rangeet Pan and Hridesh Rajan show how to decompose a trained CNN into modules, each responsible for one output class, so that a module can be reused in a new model or used to replace a faulty part of an existing one. Building a network by composing modules avoids retraining, and it makes a model repairable, because the part responsible for a problem can simply be swapped out.

This work is part of Modular and Dependable AI; it extends On Decomposing a Deep Neural Network into Modules, and the wider story is here. The full paper is available here.