Identifiable Shared Component Analysis of Unpaired Multimodal Mixtures

Authors: Subash Timilsina, Sagar Shrestha, Xiao Fu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The identifiability claims are thoroughly validated using synthetic and real-world data.
Researcher Affiliation Academia Subash Timilsina School of EECS Oregon State University Corvallis, OR 97331 timilsis@oregonstate.edu Sagar Shrestha School of EECS Oregon State University Corvallis, OR 97331 shressag@oregonstate.edu Xiao Fu School of EECS Oregon State University Corvallis, OR 97331 xiao.fu@oregonstate.edu
Pseudocode No The paper describes the proposed approach and formulations in text and mathematical equations, such as Equation (6) and (7), and provides algorithmic steps implicitly through description. However, there are no explicitly labeled 'Algorithm' or 'Pseudocode' blocks or figures.
Open Source Code Yes Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Yes the code is provided in the supplemental material.
Open Datasets Yes Dataset: We use two standard benchmarks of DA, i.e., Office-31 [58] and Office-Home [59]. ... Dataset: We use human lung adenocarcinoma A549 cells data from [66]. ... Dataset: We use the word embeddings from the MUSE dataset (https://github.com/ facebookresearch/MUSE) [21].
Dataset Splits Yes Each data set is split into 1534 training samples and 340 testing samples as in [27].
Hardware Specification Yes All the experiments were run on Nvidia H100 GPU.
Software Dependencies No The paper mentions software components like 'Adam optimizer [70]' and baselines such as 'DANN [25], MDD [60], MCC [61], SDAT [62], and ELS [63]' along with their GitHub sources. However, it does not specify version numbers for general programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or specific library dependencies (e.g., scikit-learn, NumPy) with their versions.
Experiment Setup Yes We set the initial learning rate of matrix and discriminator to be 0.009 and 0.00008 respectively. We set the λ = 0.1 in (7) to enforce (6c). For weak supervision experiment in F, we set β = 0.01 in (9). We generate total of 100,000 samples in each domain. For our experiment we set the batch size to be 1,000 and run (7) for 50 epochs. Our discriminator is a 6-layer multilayer perceptron (MLP) with hidden units { 1024, 521, 512, 256, 128, 64 } in each layer. All the layers use leaky Re LU activation functions [71] with a slope of 0.2 except for the last layer which has sigmoid activations. We include a label smoothing coefficient of 0.2 in the discriminator predictions as suggested in [40].