Knowledge Transfer with Interactive Learning of Semantic Relationships

Authors: Jonghyun Choi, Sung Ju Hwang, Leonid Sigal, Larry Davis

AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the proposed model in a few-shot multi-class classification scenario, where we measure classification performance on a set of target classes, with few training instances, by leveraging and transferring knowledge from anchor classes, that contain larger set of labeled instances.
Researcher Affiliation Collaboration Jonghyun Choi University of Maryland College Park, MD jhchoi@umiacs.umd.edu Sung Ju Hwang UNIST Ulsan, Korea sjhwang@unist.ac.kr Leonid Sigal Disney Research Pittsburgh, PA lsigal@disneyresearch.com Larry S. Davis University of Maryland College Park, MD lsd@umiacs.umd.edu
Pseudocode Yes We summarize the overall procedure in Algorithm 1 and describe the steps in the following subsections.
Open Source Code No The paper does not provide concrete access to source code for the methodology.
Open Datasets Yes We use two object categorization datasets: 1) Animals with Attributes (AWA) (Lampert, Nickisch, and Harmeling 2009), which consists of 50 animal classes and 30,475 images, 2) Image Net-50 (Hwang, Grauman, and Sha 2013), which consists of 70,380 images of 50 categories.
Dataset Splits Yes For testing and validation set, we use a 50/50 split of the remaining samples, excluding the training samples.
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes We evaluate the performance of knowledge transfer by measuring the classification accuracy of each model on the target classes in a challenging set-up that has only a few training samples (2, 5 and 10 samples per class, few-shot learning) with a prior learned with anchor classes that have a larger numbers of training samples (30 samples per class).