Learning by Transferring from Unsupervised Universal Sources

Authors: Yu-Xiong Wang, Martial Hebert

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present experimental results evaluating our unsupervised sources (UUS) as well as our HTL approach (MT-SVM) on standard recognition benchmarks, comparing several state-of-the-art methods, and validating across tasks and categories the generality of our sources.
Researcher Affiliation Academia Yu-Xiong Wang and Martial Hebert Robotics Institute, Carnegie Mellon University {yuxiongw, hebert}@cs.cmu.edu
Pseudocode No The paper describes algorithms verbally and with mathematical formulations (Eqn. 1, 2, 4, 5, 6) but does not present a formal pseudocode block or algorithm listing with structured steps.
Open Source Code No The paper does not provide a statement about releasing source code or a link to a code repository for the methodology described.
Open Datasets Yes Here, for purpose of reproducibility, we simply use the ILSVRC 2012 training dataset without access to the label information, leading to N = 1.2M unlabeled images D. ... We use Webcam as the target domain... We view the ILSVRC 2012 training dataset as the source domain... The Office dataset contains 31 classes... SUN-397 dataset (Xiao et al. 2014)... UUSs generated on PASCAL 2007.
Dataset Splits Yes Subset A: we focus on the 16 common classes between Webcam and ILSVRC as our target categories... 1 labeled training and 10 testing images per category are randomly selected on the Webcam domain, i.e., one-shot transfer and a balanced test set across categories. Therefore, each test split has 160 examples.
Hardware Specification No The paper mentions using 'convolutional neural network (CNN) features pre-trained on ILSVRC 2012' and extracting a 'd = 4,096-D feature vector fc7', but it does not specify any hardware details like GPU models, CPU models, or memory used for their own model training or experimentation.
Software Dependencies No The paper mentions tools like 'convolutional neural network (CNN) features', 'SVM', 'elastic net regularization', and 'feature-sign search', but it does not provide specific version numbers for any software or libraries used in their implementation.
Experiment Setup Yes Using an augmented pseudo-labeled dataset DAUG of C =30 pseudo-classes with K+G=6+50 samples per pseudo-class, we generate S =10 split PBCs. Repeating T =2,000 subsampling in parallel, we have generated J =20K source hypotheses in total. ... For all our experiments, we then fixed α = 10, and tuned γ to minimize the leave-one-out-error.