Supervised Transfer Sparse Coding

Abstract

A combination of the sparse coding and transfer learning techniques was shown to be accurate and robust in classification tasks where training and testing objects have a shared feature space but are sampled from different underlying distributions, i.e., belong to different domains. The key assumption in such case is that in spite of the domain disparity, samples from different domains share some common hidden factors. Previous methods often assumed that all the objects in the target domain are unlabeled, and thus the training set solely comprised objects from the source domain. However, in real world applications, the target domain often has some labeled objects, or one can always manually label a small number of them. In this paper, we explore such possibility and show how a small number of labeled data in the target domain can significantly leverage classification accuracy of the state-of-the-art transfer sparse coding methods. We further propose a unified framework named supervised transfer sparse coding (STSC) which simultaneously optimizes sparse representation, domain transfer and classification. Experimental results on three applications demonstrate that a little manual labeling and then learning the model in a supervised fashion can significantly improve classification accuracy.

Transfer learning for estimating occupancy and recognizing activities in smart buildings

“…Transferring knowledge from the source gives acceptable idea about the target without the need of instances from the target. This method takes into consideration the assumption of the importance of a few target data in the training [10,37]. The obtained result without adding target data in the training is 30%.…”

“…To discuss the proposed method, we split it into two parts: sparse coding and domain transfer. The algorithm was inspired from [37].…”

“…For this problem, the training dataset is a combination of a huge number of instances from the source domain and few instances from the target domain. Indeed, it has been proven by [37] that a small number of samples from the target domain added to the training can improve significantly the generalization of the created model. For the evaluation, the testing dataset is a combination of source and target instances as shown in Figure 9.…”

Little Is Much: Bridging Cross-Platform Behaviors through Overlapped Crowds

“…Recently, a number of approaches have studied how to conduct semi-supervised learning using non-negative matrix factorization (Koren 2008) (Singh and Gordon 2008) (Jiang et al 2014)(Al-Shedivat et al 2014. For multi-label learning, Liu et al proposed constrained NMF to minimize the difference between input patterns and class memberships (Liu, Jin, and Yang 2006).…”

Dictionary Learning with Mutually Reinforcing Group-Graph Structures

“…Following this way, many variants of KSVD appear, such as DKSVD (Zhang and Li 2010) and LCKSVD (Jiang, Lin, and Davis 2011). More recently, several methods combine sparse coding with transfer learning (Huang et al 2013;Al-Shedivat et al 2014), which ensure the learned dictionary to be more suitable for testing data. For learning more compact dictionary, group sparse representation is proposed in (Bengio et al 2009;Gao et al 2010;Chi et al 2013).…”