With transfer learning, one set of tasks is used to bias learning and improve performance on another task. However, transfer learning may actually hinder performance if the tasks are too dissimilar. As described in this paper, one challenge for transfer learning research is to develop approaches that detect and avoid negative transfer using very little data from the target task. 1

Abstract. In this paper, we propose a novel graph-based method for knowledge transfer. We model the transfer relationships between source tasks by embedding the set of learned source models in a graph using transferability as the metric. Transfer to a new problem proceeds by mapping the problem into the graph, then learning a function on this graph that automatically determines the parameters to transfer to the new learning task. This method is analogous to inductive transfer along a manifold that captures the transfer relationships between the tasks. We demonstrate improved transfer performance using this method against existing approaches in several real-world domains. 1

There has been little research into how end users might be able to communicate advice to machine learning systems. If this resource—the users themselves—could somehow work hand-in-hand with machine learning systems, the accuracy of learning systems could be improved and the users ’ understanding and trust of the system could improve as well. We conducted a think-aloud study to see how willing users were to provide feedback and to understand what kinds of feedback users could give. Users were shown explanations of machine learning predictions and asked to provide feedback to improve the predictions. We found that users had no difficulty providing generous amounts of feedback. The kinds of feedback ranged from suggestions for reweighting of features to proposals for new features, feature combinations, relational features, and wholesale changes to the learning algorithm. The results show that user feedback has the potential to significantly improve machine learning systems, but that learning algorithms need to be extended in several ways to be able to assimilate this feedback. ACM Classification: H.5.2 [Information interfaces and presentation (e.g., HCI)] User Interfaces: Theory and methods, Evaluation/methodology. H.1.2 [Models and Principles]: User/Machine Systems: Human information processing,

Abstract. We introduce a problem called relevant subtask learning, a variant of multi-task learning. The goal is to build a classifier for a taskof-interest having too little data. We additionally have data for other tasks but only some are relevant, meaning that they contain samples classified in the same way as in the task-of-interest. The problem is how to utilize this “background data ” to improve the classifier in the task-ofinterest. We show how to solve the problem for logistic regression classifiers, and demonstrate that the solution works better than a comparable multi-task learning model. The key is to assume that data of all tasks are mixtures of relevant and irrelevant samples, and model the irrelevant part with a sufficiently flexible model such that it does not distort the model of relevant data. Key words: multi-task learning, partially relevant data, relevant subtask learning, transfer learning 1

Many learning tasks in adversarial domains tend to be highly dependent on the opponent. Predefined strategies optimized for play against a specific opponent are not likely to succeed when employed against another opponent. Learning a strategy for each new opponent from scratch, though, is inefficient as one is likely to encounter the same or similar opponents again. We call this particular variant of inductive transfer a concept recall problem. We present an extension to AdaBoost called ExpBoost that is especially designed for such a sequential learning tasks. It automatically balances between an ensemble of experts each trained on one known opponent and learning the concept of the new opponent. We present and compare results of Exp-Boost and other algorithms on both synthetic data and in a simulated robot soccer task. ExpBoost can rapidly adjust to new concepts and achieve performance comparable to a classifier trained exclusively on a particular opponent with far more data.

Abstract. Transfer learning is the improvement of learning in a new task through the transfer of knowledge from a related task that has already been learned. While most machine learning algorithms are designed to address single tasks, the development of algorithms that facilitate transfer learning is a topic of ongoing interest in the machine-learning community. This chapter provides an introduction to the goals, formulations, and challenges of transfer learning. It surveys current research in this area, giving an overview of the state of the art and outlining the open problems. The survey covers transfer in both inductive learning and reinforcement learning, and discusses the issues of negative transfer and task mapping in depth.

There is often a need to adapt supervised classifiers such as semantic concept detectors across different domains of data. This paper describes a generic framework for function-level classifier adaptation based on regularized loss minimization. It directly modifies the decision function of an existing classifier of any type into a classifier for a new domain, based on limited labeled data in the new domain and no “old data”, which makes it an efficient and flexible framework. We then extend this framework to adapt multiple classifiers into one classifier, with the weights of existing classifiers learned automatically to reflect their utility. We elaborate on two concrete adaptation algorithms derived from the framework, namely adaptive SVM and multi-adaptive SVM, for one-toone and many-to-one adaptation respectively. In the experiments of adapting semantic concept detectors across video channels/types, our adaptation approach is proven to be superior to using original (unadapted) classifiers or building new ones in terms of accuracy and labeling effort.

In multimedia, machine learning techniques are often applied to build models to map low-level feature vectors into semantic labels. As data such as images and videos come from a variety of domains (e.g., genres, sources) with different distributions, there is a benefit of adapting models trained from one domain to other domains in terms of improving performance and reducing computational and human cost. In this thesis, we focus on a generic adaptation setting in multimedia, where supervised classifiers trained from one or more auxiliary domains are adapted to a new classifier that works well on a target domain with limited labeled examples. Our main contribution is a discriminative framework for function-level classifier adaptation based on regularized loss minimization, which adapts classifiers of any type by modifying their decision functions in an efficient and principled way. Two adaptation algorithms derived from this general framework, adaptive support vector machines (aSVM) and adaptive kernel logistic regression (aKLR), are discussed in details. We further extend this framework by integrating domain analysis approaches that measure and weight the utility of auxiliary

1.2 Related Work.............................................. 3

Statistical relational learning (SRL) algorithms combine ideas from rich knowledge representations, such as first-order logic, with those from probabilistic graphical models, such as Markov networks, to address the problem of learning from multi-relational data. One challenge posed by such data is that individual instances are frequently very large and include complex relationships among the entities. Moreover, because separate instances do not follow the same structure and contain varying numbers of entities, they cannot be effectively represented as a feature-vector. SRL models and algorithms have been successfully applied to a wide variety of domains such as social network analysis, biological data analysis, and planning, among others. Markov logic networks (MLNs) are a recently-developed SRL model that consists of weighted first-order clauses. MLNs can be viewed as templates that define Markov networks when provided with the set of constants present in a domain. MLNs are therefore very powerful because they inherit the expressivity of first-order logic. At the same time, MLNs can flexibly deal with noisy or uncertain data to produce probabilistic predictions for a set of propositions. MLNs have also been shown to subsume several other popular SRL models. The expressive power of MLNs comes at a cost: structure learning, or learning the first-order clauses