RegBN: Batch Normalization of Multimodal Data with Regularization

Authors: Morteza Ghahremani Boozandani, Christian Wachinger

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate the effectiveness of Reg BN on eight databases from five research areas, encompassing diverse modalities such as language, audio, image, video, depth, tabular, and 3D MRI. The proposed method demonstrates broad applicability across different architectures such as multilayer perceptrons, convolutional neural networks, and vision transformers, enabling effective normalization of both lowand high-level features in multimodal neural networks.
Researcher Affiliation Academia Morteza Ghahremani1,2 Christian Wachinger1,2 1Lab for AI in Medical Imaging (AI-Med), Department of Radiology, Technical University of Munich (TUM), Germany 2Munich Center for Machine Learning (MCML), Germany {morteza.ghahremani, christian.wachinger}@tum.de
Pseudocode Yes Algorithm 1 Pseudocode of Reg BN
Open Source Code Yes Our code is openly available at https://mogvision.github.io/Reg BN.
Open Datasets Yes The ADNI dataset [19] includes 3D MRI scans of patients along with rich clinical information organized in a low-dimensional tabular format (see Appendix ??). IEMOCAP [8], CMU-MOSI, and CMU-MOSEI [63] datasets.
Dataset Splits Yes Area Dataset Modalities Samples (#) Prediction task Training Test (Table 1 lists training and test sample counts for each dataset). Similar to [60], we partitioned the data into five separate and non-overlapping folds, ensuring that each fold had a balanced distribution of diagnosis, age, and sex.
Hardware Specification Yes We employed an NVIDIA GTX 1080 Ti with 12GB VRAM for image experiments and an NVIDIA A100 with 80GB VRAM for video experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as Python library versions or framework versions, which would be needed for replication.
Experiment Setup Yes The default parameters and settings for Reg BN are reported in Appendix ??. Experimental details are provided in Appendix ??.