Hypothesis Disparity Regularized Mutual Information Maximization
Authors: Qicheng Lao, Xiang Jiang, Mohammad Havaei8243-8251
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we evaluate the proposed HDMI approach on three benchmark datasets for domain adaptation in the context of unsupervised hypothesis transfer. We show that (i) the proposed HD regularization is effective in minimizing the undesirable disagreements among different target hypotheses and stabilizing the MI maximization process; (ii) Compared to direct MI maximization with single hypothesis or multiple hypotheses, the HD regularization facilitates the positive transfer of multiple modes from source hypotheses, and as a result, the target hypotheses obtained by HDMI preserve more transferable knowledge from each source hypothesis; (iii) HDMI uses well-calibrated predictive uncertainty to achieve effective MI maximization; and (iv) HDMI works through learning better representations shared by different target hypotheses. Overall, HDMI achieves new state-of-the-art performance in unsupervised hypothesis transfer learning. |
| Researcher Affiliation | Collaboration | 1 West China Biomedical Big Data Center, West China Hospital of Sichuan University, Chengdu, China 2 MILA, Universit e de Montr eal, Montreal, Canada 3 Imagia, Montreal, Canada |
| Pseudocode | No | No structured pseudocode or algorithm blocks were found in the provided text. |
| Open Source Code | No | The paper mentions "A US provisional patent application has been filed for protecting at least one part of the innovation disclosed in this article." and does not provide any statement or link about the availability of open-source code for the described methodology. |
| Open Datasets | Yes | Datesets Office-31 (Saenko et al. 2010) has three domains: Amazon (A), DSLR (D) and Webcam (W), with 31 classes and 4,652 images. Office-Home (Venkateswara et al. 2017) is a more challenging dataset with 65 classes and 15,500 images in four domains: Artistic images (Ar), Clip art (Cl), Product images (Pr) and Real-World images (Rw). Vis DAC (Peng et al. 2018) is a large-scale dataset with 12 classes, with 152,397 Synthetic images in the source domain and 55,388 Real images in the target domain. |
| Dataset Splits | No | The paper mentions using training data for learning source hypotheses and target data for adaptation, but does not provide explicit percentages, sample counts, or references to predefined train/validation/test splits for reproducibility. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) needed to replicate the experiment. |
| Experiment Setup | Yes | Note that we set the number of hypotheses M = 2 by default unless otherwise stated, since empirical results suggest that the HD regularization between two hypotheses suffices HDMI for better performance. ... We validate the robustness of HDMI in terms of the number of hypotheses M and the hyperparameter λ, we perform experiments on A D, Office31 with different configurations of M and λ, and summarize the results in Table 4. |