An Efficient Approach to Informative Feature Extraction from Multimodal Data

Authors: Lichen Wang, Jiaxiang Wu, Shao-Lun Huang, Lizhong Zheng, Xiangxiang Xu, Lin Zhang, Junzhou Huang5281-5288

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluation implies that our approach learns more informative feature mappings and is more efficient to optimize. In this section, we evaluate Soft-HGR in the following aspects: To verify the relationship between the HGR features and Soft-HGR feature is linear; To compare the efficiency and numerical stability of CCA based models and Soft-HGR; To demonstrate the power of semi-supervised Soft-HGR on discriminative tasks with limited labels; To show the performance of Soft-HGR on more than two modalities and missing modalities.
Researcher Affiliation Collaboration Lichen Wang,1 Jiaxiang Wu,2 Shao-Lun Huang,1 Lizhong Zheng,3 Xiangxiang Xu,4 Lin Zhang,1 Junzhou Huang5 1Tsinghua-Berkeley Shenzhen Institute, Tsinghua University, 2Tencent AI Lab 3Department of EECS, Massachusetts Institute of Technology 4Department of Electronic Engineering, Tsinghua University 5Department of CSE, The University of Texas at Arlington
Pseudocode Yes Algorithm 1 Evaluate Soft-HGR on a mini-batch
Open Source Code No The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes MNIST handwritten image dataset (Le Cun et al. 1998), which consists of 60K/10K gray-scale digit images of size 28 28 as training/testing sets. University of Wisconsin X-ray Microbeam Database (XRMB) (Westbury 1994). KKBox s Music Recommendation Dataset (Chen et al. 2018).
Dataset Splits Yes The hyper-parameters for each model are determined by their best average performance on validation set on 5-fold cross validation. we use the last 20% of 7M training data as test set.
Hardware Specification Yes Both optimizations are executed on a Nvidia Tesla K80 GPU with mini-batch SGD of 5K batchsize.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries, such as Python versions or deep learning framework versions.
Experiment Setup Yes Both optimizations are executed on a Nvidia Tesla K80 GPU with mini-batch SGD of 5K batchsize. The DNN structure for each neural branches may vary, depending on the statistical property of the inputs. The output feature dimensions k is chosen to be 80 for all methods. The hyper-parameters for each model are determined by their best average performance on validation set on 5-fold cross validation. batch normalization (Ioffe and Szegedy 2015) is applied before Re LU activation function to ensure better convergence.