A Unified Perspective on Multi-Domain and Multi-Task Learning

Authors: Yongxin Yang and Timothy Hospedales

ICLR 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments across this range of problems demonstrate that our framework outperforms a variety of alternatives. (Abstract) and We demonstrate our framework on five experimental settings: MDL, ZSDA, MTL, ZSL and MDMT. (Section 4)
Researcher Affiliation Academia Yongxin Yang & Timothy M. Hospedales Electronic Engineering and Computer Science Queen Mary, University of London {yongxin.yang, t.hospedales}@qmul.ac.uk
Pseudocode No The paper describes the model and framework but does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide a specific link or explicit statement about the availability of the source code for the methodology described.
Open Datasets Yes This classic dataset3 collects exam grades of 15,362 students from 139 schools. 3Available at http://multilevel.ioe.ac.uk/intro/datasets.html (Section 4.1) and Animal with Attributes (Lampert et al., 2009) (Section 4.3) and The restaurant & Consumer Dataset, introduced by Vargas-Govea et al. (2011) (Section 4.4).
Dataset Splits No The paper specifies training/test splits (e.g., 'In each case the training/test split is 50%/50%') but does not explicitly mention a separate validation set or split ratio for validation data.
Hardware Specification No The paper mentions 'NVIDIA Corporation for the donation of the GPUs used for this research' in the acknowledgements, but does not specify exact GPU models or any other hardware details (CPU, memory, specific cloud instances).
Software Dependencies No The paper mentions using 'Caffe framework (Jia et al., 2014)' but does not provide specific version numbers for Caffe or any other software dependencies.
Experiment Setup Yes Preliminary experiments show K = D log(D) leads to satisfactory solutions for all datasets. Though we don t place regularisation terms on P or Q, a non-linear function σ(x) = max(0, x) (i.e., Re LU activation function) is placed to encourage sparse models g Q(z(i)) = σ(z(i)Q). The choice of loss function for regression and classification is Euclidean loss and Hinge loss respectively. (Section 4)