AccGenSVM: Selectively Transferring from Previous Hypotheses

Authors: Diana Benavides-Prado, Yun Sing Koh, Patricia Riddle

IJCAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here, we present experiments and discuss results for binary classification using Acc Gen SVM and other publicly available HTL methods for homogeneous transfer.
Researcher Affiliation Academia Diana Benavides-Prado Dept. of Computer Science The University of Auckland dben652@aucklanduni.ac.nz; Yun Sing Koh Dept. of Computer Science The University of Auckland ykoh@cs.auckland.ac.nz; Patricia Riddle Dept. of Computer Science The University of Auckland pat@cs.auckland.ac.nz
Pseudocode Yes Algorithm 1: Pseudo-code for transfer with Acc Gen SVM
Open Source Code Yes Acc Gen SVM1 is built on top of Lib SVM [Chang and Lin, 2011], using available KL divergence [Hausser and Strimmer, 2014] and FNN implementations [Beygelzimer et al., 2013]. 1Software available at: https://github.com/nanarosebp/ Ph DProject/tree/master/Acc Gen SVM
Open Datasets Yes Caltech256 is a benchmark dataset on image recognition... Image Net is a large benchmark dataset... Office is a small dataset... Caltech-256 object category dataset. [Griffin et al., 2007]
Dataset Splits No The paper mentions training and test splits, but no explicit validation set split is described. 'Transfer level: we extract random samples of sizes 10%, 20% and 30% for training. ... We also select independent random binary samples of size 20%, 30%, and 50% as test sets.'
Hardware Specification No No specific hardware details (such as GPU/CPU models or memory specifications) used for running experiments are provided.
Software Dependencies Yes Acc Gen SVM1 is built on top of Lib SVM [Chang and Lin, 2011], using available KL divergence [Hausser and Strimmer, 2014] and FNN implementations [Beygelzimer et al., 2013]. For FNN, 'R package version 1.1, 2013.' is cited. For KL divergence, 'R package v.1.2.1.' is cited.
Experiment Setup Yes Based on a sensitivity analysis of the regularization parameter C and the γ parameter for the RBF kernel, we set C = 1 and γ = 1/f, with f number of features. We use a KL divergence threshold of 0.3 for all datasets. For FNN, we work with 3 nearest neighbours.