Transfer Learning via Learning to Transfer

Authors: Ying WEI, Yu Zhang, Junzhou Huang, Qiang Yang

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also theoretically analyse the algorithmic stability and generalization bound of L2T, and empirically demonstrate its superiority over several state-of-the-art transfer learning algorithms.Datasets We evaluate the L2T framework on two image datasets, Caltech-256 (Griffin et al., 2007) and Sketches (Eitz et al., 2012).We compare L2T with the following nine baseline algorithms in three classes:Non-transfer: Original builds a model using labeled data in a target domain only.Common latent space based transfer learning algorithms: TCA (Pan et al., 2011), ITL (Shi & Sha, 2012), CMF (Long et al., 2014), LSDT (Zhang et al., 2016), STL (Raina et al., 2007), DIP (Baktashmotlagh et al., 2013) and SIE (Baktashmotlagh et al., 2014).Manifold ensemble based algorithms: GFK (Gong et al., 2012).The eight feature-based transfer learning algorithms also constitute the base set A. Based on feature representations obtained by different algorithms, we use the nearestneighbor classifier to perform three-class classification for the target domain.
Researcher Affiliation Collaboration 1Hong Kong University of Science and Technology, Hong Kong 2Tencent AI Lab, Shenzhen, China.
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. It describes the framework and optimization problems mathematically.
Open Source Code No The paper does not contain any statement about releasing source code for the described methodology.
Open Datasets Yes Datasets We evaluate the L2T framework on two image datasets, Caltech-256 (Griffin et al., 2007) and Sketches (Eitz et al., 2012).
Dataset Splits Yes In total, we generate 1,000 training pairs for preparing transfer learning experiences, 500 validation pairs to determine hyperparameters of the reflection function, and 500 testing pairs to evaluate the reflection function.
Hardware Specification No The paper does not provide any specific details about the hardware used for running the experiments (e.g., CPU, GPU models, memory).
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes In building the reflection function, we use 33 RBF kernels with the bandwidth δk in the range of [2 8η : 20.5η : 28η] where η = 1 nsente Ne Ne e=1 ns e,nt e i,j=1 xs ei W xt ej W 2 2 follows the median trick (Gretton et al., 2012a).