Minimal-Entropy Correlation Alignment for Unsupervised Deep Domain Adaptation
Authors: Pietro Morerio, Jacopo Cavazza, Vittorio Murino
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide extensive experiments to assess the superiority of our framework on standard domain and modality adaptation benchmarks. ... Through an extensive experimental analysis on publicly available benchmarks for transfer object categorization, we certify the effectiveness of the proposed approach in terms of systematic improvements over former alignment methods and state-of-the-art techniques for unsupervised domain adaptation in general. |
| Researcher Affiliation | Collaboration | Pietro Morerio1, Jacopo Cavazza1 & Vittorio Murino1,2 1 Pattern Analysis and Computer Vision (PAVIS), Istituto Italiano di Tecnologia Genova, Italy 2 University of Verona, Department of Computer Science Verona, Italy |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code available at https://github.com/pmorerio/minimal-entropy-correlation-alignment |
| Open Datasets | Yes | We use digits from SVHN [Netzer et al. (2011)] as source and we transfer on MNIST. Similarly, we transfer from SYN DIGITS [Ganin & Lempitsky (2015)] to SVHN. For the object recognition task, we train a model to classify objects on RGB images from NYUD [Silberman et al. (2012)] dataset and we test on (different) depth images from the same visual categories. |
| Dataset Splits | Yes | We used the whole training sets of both datasets, following the usual protocol for unsupervised domain adaptation (SVHN s training set contains 73, 257 images). We also resized MNIST images to 32 32 pixels and converted SVHN to grayscale, according to the standard protocol. ... In summary, we propose the following minimization pipeline for unsupervised domain adaptation, which we name Minimal-Entropy Correlation Alignment (MECA) min θ [H(XS, ZS) + λ ℓlog(CS, CT )] subject to λ minimizes E(XT ). (9) In other words, in (9), we minimize the objective functional H(XS, ZS) + λ ℓlog(CS, CT ) by gradient descent over θ. While doing so, we can choose λ by validation, such that the network f( ; θ) is able, at the same time, to attain the minimal entropy on the target domain. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | While 'Tensor Flow TM' is mentioned, no specific version number for TensorFlow or any other software dependencies is provided. |
| Experiment Setup | Yes | The architecture is the very same employed in [Ganin & Lempitsky (2015)] with the only difference that the last fully connected layer (fc2) has only 64 units instead of 2048. ... We finetune a VGG in order to be comparable with ADDA baseline in [Tzeng et al. (2017)]. Covariance alignment occurs at fc8, which is replaced with a 64-unit layer. ... SYN DIGITS SVHN. Same as for SVHN MNIST, but fc1 has 3072 units. ... When aligning second order statistics, a hyper-parameter controls the balance between the reduction of the domain shift and the supervised classification on the source domain. |