Constructing informative priors using transfer learning

Raina, Rajat; Ng, Andrew Y.; Koller, Daphne

doi:10.1145/1143844.1143934

Cited by 210 publications

(157 citation statements)

References 11 publications

Supporting

Mentioning

153

Contrasting

Order By: Relevance

“…Other work on transfer (e.g., [Raina et al, 2006, Niculescu-Mizil and Caruana, 2007, Torrey et al, 2007, Richardson and Domingos, 2006) also share some resemblance with our work. They focus on improving the performance of learning by transferring previously acquired knowledge from another domain of interest.…”

Section: Discussionsupporting

confidence: 82%

Integrating representation learning and skill learning in a human-like intelligent agent

Matsuda

Cohen

et al. 2015

Artificial Intelligence

View full text Add to dashboard Cite

Building an intelligent agent that simulates human learning of math and science could potentially benefit both cognitive science, by contributing to the understanding of human learning, and artificial intelligence, by advancing the goal of creating human-level intelligence. However, constructing such a learning agent currently requires manual encoding of prior domain knowledge; in addition to being a poor model of human acquisition of prior knowledge, manual knowledge-encoding is both time-consuming and error-prone. Previous research has shown that one of the key factors that differentiates experts and novices is their different representations of knowledge. Experts view the world in terms of deep functional features, while novices view it in terms of shallow perceptual features. Moreover, since the performance of learning algorithms is sensitive to representation, the deep features are also important in achieving effective machine learning.In this work, we propose an efficient algorithm that acquires representation knowledge in the form of "deep features" for specific domains, and demonstrate its effectiveness in the domain of algebra as well as synthetic domains. We integrate this algorithm into a learning agent, SimStudent, which learns procedural knowledge by observing a tutor solve sample problems, and by getting feedback while actively solving problems on its own. We show that learning representations enhances the generality of the learning agent by reducing the requirements for knowledge engineering. Moreover, we propose an approach that automatically discovers student models using the extended SimStudent. By fitting the discovered model to real student learning curve data, we show that the discovered model is better or as good as human-generated models, and demonstrate how the discovered model may be used to improve a tutoring system's instructional strategy.vi

show abstract

Section: Discussionsupporting

confidence: 82%

Integrating representation learning and skill learning in a human-like intelligent agent

Matsuda

Cohen

et al. 2015

Artificial Intelligence

View full text Add to dashboard Cite

show abstract

“…A third form of transfer learning is model adaptation, where auxiliary data is used to regularize the parameters of a target model, which can be either generative [60][61][62][63][64][65] or discriminative [66][67][68][69][70]. Although this is sometimes denoted domain adaptation, the latter usually refers to methods that regularize the target feature space, rather than the models themselves.…”

Section: Related Workmentioning

confidence: 99%

Cross-modal domain adaptation for text-based regularization of image semantics in image retrieval systems

Pereira

Vasconcelos

2014

Computer Vision and Image Understanding

View full text Add to dashboard Cite

a b s t r a c tIn query-by-semantic-example image retrieval, images are ranked by similarity of semantic descriptors. These descriptors are obtained by classifying each image with respect to a pre-defined vocabulary of semantic concepts. In this work, we consider the problem of improving the accuracy of semantic descriptors through cross-modal regularization, based on auxiliary text. A cross-modal regularizer, composed of three steps, is proposed. Training images and text are first mapped to a common semantic space. A regularization operator is then learned for each concept in the semantic vocabulary. This is an operator which maps the semantic descriptors of images labeled with that concept to the descriptors of the associated texts. A convex formulation of the learning problem is introduced, enabling the efficient computation of concept-specific regularization operators. The third step is the selection of the most suitable operator for the image to regularize. This is implemented through a quantization of the semantic space, where a regularization operator is associated with each quantization cell. Overall, the proposed regularizer is a non-linear mapping, implemented as a piecewise linear transformation of the semantic image descriptors to regularize. This transformation is a form of cross-modal domain adaptation. It is shown to achieve better performance than recent proposals in the domain adaptation literature, while requiring much simpler optimization.

show abstract

“…In image, text analysis or robotics many methods have been devised for knowledge transfer. Related machine learning subjects include: learning from hints [104], lifelong learning [105], multi-task learning [106], cross-domain learning [107,108], cross-category learning [109] and selftaught learning [110]. EigenTransfer algorithm [111] tries to unify various transfer learning ideas representing the target task by a graph.…”

Section: Transfer Of Knowledgementioning

confidence: 99%

Optimal Support Features for Meta-Learning

Duch

Maszczyk

Grochowski

2011

Studies in Computational Intelligence

View full text Add to dashboard Cite

Meta-learning has many aspects, but its final goal is to discover in an automatic way many interesting models for a given data. Our early attempts in this area involved heterogeneous learning systems combined with a complexity-guided search for optimal models, performed within the framework of (dis)similarity based methods to discover "knowledge granules". This approach, inspired by neurocognitive mechanisms of information processing in the brain, is generalized here to learning based on parallel chains of transformations that extract useful information granules and use it as additional features. Various types of transformations that generate hidden features are analyzed and methods to generate them are discussed. They include restricted random projections, optimization of these features using projection pursuit methods, similarity-based and general kernel-based features, conditionally defined features, features derived from partial successes of various learning algorithms, and using the whole learning models as new features. In the enhanced feature space the goal of learning is to create image of the input data that can be directly handled by relatively simple decision processes. The focus is on hierarchical methods for generation of information, starting from new support features that are discovered by different types of data models created on similar tasks and successively building more complex features on the enhanced feature spaces. Resulting algorithms facilitate deep learning, and also enable understanding of structures present in the data by visualization of the results of data transformations and by creating logical, fuzzy and prototype-based rules based on new features. Relations to various machine-learning approaches, comparison of results, and neurocognitive inspirations for meta-learning are discussed.

show abstract

Constructing informative priors using transfer learning

Cited by 210 publications

References 11 publications

Integrating representation learning and skill learning in a human-like intelligent agent

Integrating representation learning and skill learning in a human-like intelligent agent

Cross-modal domain adaptation for text-based regularization of image semantics in image retrieval systems

Optimal Support Features for Meta-Learning

Contact Info

Product

Resources

About