Applying Product Line Engineering Concepts to Deep Neural Networks

Deep Neural Networks (DNNs) are increasingly being used as a machine learning solution thanks to the complexity of their architecture and hyperparameters-weights. A drawback is the excessive demand for massive computational power during the training process. Not only as a whole but parts of neural networks can also be in charge of certain functionalities. We present a novel challenge in an intersection between machine learning and variability management communities to reuse modules of DNNs without further training. Let us assume that we are given a DNN for image processing that recognizes cats and dogs. By extracting a part of the network, without additional training a new DNN should be divisible with the functionality of recognizing only cats. Existing research in variability management can offer a foundation for a product line of DNNs composing the reusable functionalities. An ideal solution can be evaluated based on its speed, granularity of determined functionalities, and the support for adding variability to the network. The challenge is decomposed in three subchallenges: feature extraction, feature abstraction, and the implementation of a product line of DNNs.