Machine Learning based Domain Adaptation (MLB-DA) for Multiple Source Classification and Fusion

Generalizing models learned on one domain to novel domains has been a major obstacle in the quest for universe object recognition. The performance of the learned models degrades significantly when testing on novel domains due to the presence of domain shift. In this proposal, we aim to develop a deep learning-based multi-source self-correcting approach to fuse data with different modalities at the data-level to maximize their capabilities to detect unanticipated events/targets, by leveraging our previous experience in machine learning and heterogeneous data fusion. In this proposal, we propose a Machine Learning based Domain Adaptation (MLB-DA) that leverages unsupervised data to bridge the source and target domain distributions closer in a learned joint feature space. The proposed deep neural network approach holds great capability of adapting to changes of the input distribution allowing self-correcting multiple source classification and fusion. It is focused on learning features that combine (i) discriminativeness and (ii) domain-invariance. The classifiers can adapt to the target domain with different distribution without retraining new input data.