Unsupervised domain adaptation deals with the scenario where there is scarcity of labeled data for the training of machine learning models. Traditionally, the knowledge of a learner is transferred from a source domain to a target domain in a conventional setting. In this paper, we propose a novel framework for Multi-Source Domain Adaptation which can train the model using multiple source domains and can generalize well on a target dataset. It mitigates the domain shift caused by different domains by (1) coordinating the statistical moments for all combinations of the source and target domain distributions and (2) by collaboratively extracting the deep features from the shared representation space. Extensive experiments have been performed for the task of image classification using different datasets widely varying in size and complexity in order to demonstrate the effectiveness of our approach compared to the naive source-combined domain adaptation setting.

To train a classifier on a specific domain, often called the target domain, we need labeled data. However, there might be non availability of the labeled data in this domain. In this scenario, we look for a related domain called the source domain, where availability of labeled data is abundant in number. The lack of availability of labeled data in the target domain poses a serious problem and several domain adaptation (DA) approaches have been put forward to cope up with this problem. Existing DA methods seek a subspace common between both the domains (source and target domains) where the distribution difference is minimal and perform manual parameter sensitivity tests to find apposite value of each parameter for their respective objective function. However, for distorted original data, obtaining a common subspace is a challenging task and condensing manual parameter sensitivity testing is also costly and a time-intensive process. To overcome these challenges, some DA methods consider particle swarm optimization (PSO) technique. However, none of the existing DA methods simultaneously tackle these challenges. Therefore, in this paper, we put forward a method called Feature and Parameter Selection approach for visual Domain Adaptation (FPSDA) to address these challenges. In FPSDA, a suitable subset of features across both the domains and an apposite value of each parameter are simultaneously chosen using a PSO approach. Moreover, to guide the PSO, the objective functions of Joint Geometrical and Statistical Alignment (JGSA) [1] method along with preserving original similarity of data is considered as an objective function for our proposed approach. Full Scale experiments on benchmark datasets for cross-domain adaptation verify that FPSDA performs better than many state-of-the-art classic machine learning and domain adaptation approaches.

In current scenario, transfer learning or domain adaptation has been emerged as fruitful approach to handle problems of distribution mismatch between the training and test data. Standard machine learning techniques utilize labeled data for getting better performance. In practical scenario, due to scarcity of labeled data, the classifier trained on the source domain cannot be utilized efficiently for classifying the target domain data. So, there is a need to use the previously acquired knowledge from a related source domain to classify the information in the target domain. Previous works on single/multiple source domain adaptation have achieved substantial growth in tackling this concern. In this work, we have proposed a novel unsupervised multi-source based approach MSVDA for visual domain adaptation in which data from multiple labeled source domains are utilized to classify the information in the target domain containing unlabeled data. Our proposed approach MSVDA extends the existing Maximum Mean Discrepancy criteria across various source domains and also preserves the discriminative information of various source domains. Further, it learns an optimal classification that diminishes the empirical risk and enhances the consistency rate between the prediction function and the manifold. Various experiments on the two publicly available datasets with multiple-domain scenario settings (double-domain, triple-domain and quadruple-domain scenarios) have demonstrated the efficacy of our proposed method MSVDA over other existing methods.

Domain adapted machine learning is driven by the possibilities of learning from source data distribution to understand different target data distributions. An assumption is made that one application (source) domain always has enough labeled information, but the other related application (target) may contain information that is partially labeled or completely unlabeled. Therefore, it is necessary to train the target domain classifier using enough labeled information of the source domain. However, contrary to primitive assumptions, the source domain and target domain data need not have the same distribution. Therefore, we can't directly use data of source domain to train classifier for data of target domain. Existing approaches can be deprived of one or more objectives: perform geometric diffusion on the manifold, align the cross-domain distributions, preserve the discriminative information using metric learning. Here, we have proposed a novel framework that aims to meet all such objectives. In this framework, we proposed two methods, statistical and geometrical alignment using metric learning with pseudo labels (SGA-MDAP) and without pseudo labels (SGA-MDA) in visual domain adaptation. It has been demonstrated through various experiments that our framework outperforms various state-of-the-art methods over four different real-world cross-domain visual identification datasets such as PIE face, ORL face, Yale face, and Office Caltech.

Data such as videos, genetic information, etc. from real-world applications reside in a high dimensional space. Before performing classification, it is required to project the data from the high dimensional space to a lower dimensional space without losing too much information. Linear discriminant analysis (LDA) is one of the most widely used methods for dimensionality reduction, that maximizes the ratio of the between-class scatter and total data scatter in the projected space using the labeled information. However, in the real world scenario, labeled information is hardly ever available in large quantities, but an abundant amount of unlabeled data is available. In this paper, we propose a Semi-Supervised Discriminant Analysis method called SSDARD, which considers the unlabeled information in the form of a k-NN graph. Different from the existing semi-supervised dimensionality reduction algorithms, our algorithm is more consistent in propagating the label information from labeled data to unlabeled data due to the use of relative distance function instead of normal Euclidean distance function to generate the k-NN graph. To find an appropriate relative distance function, we use pairwise constraints generated from labeled data and satisfy them using Bregman projection. Since the projection is not orthogonal, we require an appropriate subset of constraints. In order to select such subset of constraints, we have further developed a framework called MO-SSDARD, which uses an evolutionary algorithm while optimizing various cluster validity indices simultaneously. The experimental results on various datasets show that our proposed method is superior than various methods concerning various validity indices.