Learning Transferrable Representations for Unsupervised Domain Adaptation

Authors: Ozan Sener, Hyun Oh Song, Ashutosh Saxena, Silvio Savarese

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that the proposed method significantly outperforms state-of-the-art algorithms in both object recognition and digit classification experiments by a large margin.
Researcher Affiliation Collaboration Ozan Sener1, Hyun Oh Song1, Ashutosh Saxena2, Silvio Savarese1 Stanford University1 Brain of Things2 {ozan,hsong,asaxena,ssilvio}@cs.stanford.edu
Pseudocode Yes Algorithm 1 Transduction with Domain Shift
Open Source Code Yes Learned models and the source code can be reached from the project webpage http://cvgl.stanford.edu/transductive_adaptation.
Open Datasets Yes We use MNIST [19], Street View House Number [21] and the artificially generated version of MNIST -MNIST-M[11] to experiment our algorithm on the digit classification task. (...) In addition, we use the Office [25] dataset to evaluate our algorithm on the object recognition task.
Dataset Splits No The paper evaluates in a transductive setup, feeding 'training images' for source and target domains, but does not specify train/validation/test splits with percentages, counts, or predefined split citations for reproducibility of data partitioning for the model training itself.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running experiments.
Software Dependencies No The paper mentions using AlexNet and LeNet architectures and AdaGrad, but does not specify version numbers for any software dependencies or frameworks.
Experiment Setup Yes We set the feature dimension as 128. We use stochastic gradient descent to learn the feature function with Ada Grad[8]. We initialize convolutional weights with truncated normals having std-dev 0.1, biases with constant value 0.1, and use a learning rate of 2.5e-4 with batch size 512. We start the rejection penalty with γ = 0.1 and linearly increase with each epoch as γ = #epoch/M + 0.1. In our experiments, we use M = 20, λ = 0.001 and α = 1.