Transfer Learning: Problems And Methods

Theubiquitousgrowthofdatavolumesandcomputingresourceshasacceleratedma-chine learning with rapid advances in both theory and practice. The success of traditionalmachine learning algorithms often relies on the fundamental assumption that observeddata is under a stationary generating mechanism. However, in a wide range of machinelearning application domains, including big data analytics, natural language processing,computer vision, and bioinformatics, this assumption may be too restricted to be satisfied.Therefore, analyzing and mining the massive data under non-stationary environments isamong the greatest challenges of modern machine learning. Transfer learning relaxes theassumption of traditional machine learning that the training data and testing data shouldbe sampledindependentlyfrom anidentical probabilitydistribution, thus itcanbe appliedto discover domain-invariant intrinsic features and structures underlying two different butrelated domains, which establishes successful transfer and reutilization of supervised in-formationacrossdomains. Asoneofthebasictoolsforaddressingthelearningtaskwhichmay fail with scarcity of labeled data, transfer learning remains an open paradigm withseveral unsolved challenges. To boost cross-domain classification and prediction tasks,thisthesispresentsasystematicstudyontheopenissuesandsolutionsoftransferlearning.Transfer learning involves several critical issues and challenges: overfitting, under-fitting, under-adaptation, and negative-transfer. Overfitting and underfitting may hap-pen when modeling the unknown probability distribution based on observed data; Under-adaptationandnegative-transfer mayhappenwhen adaptingtheunknownprobabilitydis-tributions across domains: under-adaptation refers to the condition that the distributionmismatch cannot be corrected sufficiently; negative-transfer refers to the condition thatthe auxiliary task deteriorates the target task unintentionally. This thesis addresses theunderfitting, under-adaptation, and negative-transfer issues, analyzes the intrinsic causes,anddesignsspecificlearningmodels. Thenovelcontributionsaresummarizedasfollows.1. For addressing the negative-transfer problem, a graph regularized collective ma-trix factorization model is proposed to1) construct semantic features for cross-domainknowledge transfer,2) enhance the transferability of semantic features, and3) combat thenegativeeffectsofsemanticfeatures. Thismodelsynthesizestheadvantagesoftwomain-stream methods to establish effective tradeoff between under-transfer and over-transfer. 2. For addressing the under-adaptation problem, a joint adaptation regularizationframework is proposed, which extends the maximum mean discrepancy to measure thedivergence between different joint probability distributions; Both feature learning andclassifier learning are explored to adapt the mismatched joint probability distributionsacross domains, from which three transfer learning methods based on linear regression,support vector machines, and principal component analysis are formulated; The general-ization error bound of these methods are theoretically analyzed via the statistical learningtheory. Furthermore, a novel criterion based on the low-rank approximation error of ker-nelmatrixisproposedforcomparingdifferentprobabilitydistributionsandadaptingthemsufficiently across domains, with theoretical analysis on the approximation error bound.3. Foraddressingtheunderfitting,under-adaptation,andnegative-transferproblems,deep learning is explored to extend the maximum mean discrepancy to nonlinear distribu-tion discrepancy, and a unified robust deep representation adaptation model is developedto tackle the three problems collaboratively. Finally, a transfer cross-validation strategyis proposed for model selection of unsupervised transfer learning without target labels.