Multi-Domain Adversarial Learning

Authors: Alice Schoenauer-Sebag, Louise Heinrich, Marc Schoenauer, Michele Sebag, Lani F. Wu, Steve J. Altschuler

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This paper presents a multi-domain adversarial learning approach, MULANN, to leverage multiple datasets with overlapping but distinct class sets, in a semisupervised setting. Our contributions include: i) a bound on the averageand worst-domain risk in MDL, obtained using the H-divergence; ii) a new loss to accommodate semi-supervised multi-domain learning and domain adaptation; iii) the experimental validation of the approach, improving on the state of the art on three standard image benchmarks, and a novel bioimage dataset, CELL.
Researcher Affiliation Academia Alice Schoenauer Sebag1, alice.schoenauer@polytechnique.org Louise Heinrich1, louise.heinrich@ucsf.edu Marc Schoenauer2, marc.schoenauer@inria.fr Michele Sebag2, sebag@lri.fr Lani F. Wu1, lani.wu@ucsf.edu Steven J. Altschuler1 steven.altschuler@ucsf.edu 1 Department of Pharmaceutical Chemistry UCSF, San Francisco, CA 94158 2 INRIA-CNRS-UPSud-UPSaclay TAU, U. Paris-Sud, 91405 Orsay
Pseudocode No The paper describes the architecture and loss function but does not provide pseudocode or a clearly labeled algorithm block.
Open Source Code Yes Code and data: github.com/Altschuler Wu-Lab/Mu LANN
Open Datasets Yes The DA setting considers three benchmarks: DIGITS, including the well-known MNIST and MNIST-M (Le Cun et al., 1998; Ganin et al., 2016); Synthetic road signs and German traffic sign benchmark (Chigorin et al., 2012; Stallkamp et al., 2012) and OFFICE (Saenko et al., 2010). The MDL setting considers the new CELL benchmark, which is made of fluorescence microscopy images of cells (detailed in Appendix C). Code and data: github.com/Altschuler Wu-Lab/Mu LANN
Dataset Splits Yes For OFFICE and CELL, we follow the experimental settings from Saenko et al. (2010). A fixed number of labeled images per class is used for one of the domains in all cases (20 for Amazon, 8 for DSLR and Webcam, 10 in CELL). For the other domain, 10 labeled images per class are used for half of the classes (15 for OFFICE, 4 for CELL). For DIGITS and Road Signs, all labeled source train data is used, whereas labeled target data is used for half of the classes only (5 for DIGITS, 22 for Road Signs). As in (Ganin et al., 2016), no hyper-parameter grid-search is performed for OFFICE results double cross-validation is used for all other benchmarks.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU, GPU models, memory) used for the experiments. It only mentions 'access to their GPUs' in the acknowledgements without specifying models.
Software Dependencies No The paper mentions software like 'Caffe (Jia et al., 2014)', 'Torch (Collobert et al., 2011)', and 'Sci Py (Jones et al., 2001)' along with their reference papers, but it does not specify exact version numbers for these software packages or libraries, which are necessary for full reproducibility.
Experiment Setup Yes For DANN, MADA and MULANN, the same pre-trained VGG-16 architecture (Simonyan & Zisserman, 2014) from Caffe (Jia et al., 2014) is used for OFFICE and CELL2; the same small convolutional network as Ganin et al. (2016) is used for DIGITS (see Appendix D.1 for details). The models are trained in Torch (Collobert et al., 2011) using stochastic gradient descent with momentum (ρ = 0.9). Hyper-parameter ranges can be found in Appendix D.2. Parameter DIGITS and Signs CELL Learning rate (lr) 10 3, 10 4 10 4 (+ 10 5 for 3-dom.) Lr schedule Exponentially decreasing, constant λ 0.1, 0.8 λ schedule Exponentially increasing, constant ζ 0.1, 0.8