Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Meta-Learning Approach for Joint Multimodal Signals with Multimodal Iterative Adaptation
Authors: Sehun Lee, Wonkwang Lee, Gunhee Kim
TMLR 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive evaluation in various real-world multimodal signal regression setups shows that MIA outperforms existing baselines in both generalization and memorization performances. Our code is available at https://github.com/yhytoto12/MIA. ... 5 Experiments We demonstrate our proposed MIA improves performance of state-of-the-art meta-learned INR models. |
| Researcher Affiliation | Academia | Sehun Lee EMAIL Department of Computer Science and Engineering Seoul National University Wonkwang Lee EMAIL Department of Computer Science and Engineering Seoul National University Gunhee Kim EMAIL Department of Computer Science and Engineering Seoul National University |
| Pseudocode | Yes | We present a schematic illustration in Figure 2 and provide a meta-learning algorithm of our MIA in Algorithm 1. ... We also attach Py Torch-style pseudo code for MIA in Listing 1 and SFTs in Listing 2, respectively. |
| Open Source Code | Yes | Our code is available at https://github.com/yhytoto12/MIA. |
| Open Datasets | Yes | Datasets and Metrics. Our method is evaluated across four datasets: (1) multimodal 1D synthetic functions (Kim et al., 2022a), (2) multimodal 2D Celeb A images (Xia et al., 2021), (3) ERA5 global climate dataset (Hersbach et al., 2019), and (4) Audiovisual-MNIST (AV-MNIST) dataset (Vielzeuf et al., 2018). |
| Dataset Splits | Yes | For training, we construct the support set Dtrain nm by sampling Pnm Rnm coordinate-feature pairs from each full dataset Dnm. The sampling ratio Rnm varies within predefined ranges [Rmin m , Rmax m ], independently drawn for each signal. The full dataset serves as the query set Dval nm during validation to assess both memorization and generalization capabilities. ... Table 4: List of common configurations for each dataset. ... sampling ratio [Rmin, Rmax] [0.01, 0.1] [0.001, 1] [0.001, 1] Images [0.001, 1] Audios [0.250, 1] |
| Hardware Specification | No | The paper discusses computational overheads including memory consumption and training time in Section 6.3 and Table 3. It mentions 'encountered out-of-memory (OOM) issues' but does not specify any particular GPU or CPU models, memory sizes, or other detailed hardware specifications used for the experiments. |
| Software Dependencies | No | The paper mentions 'Adam optimizer (Kingma & Ba, 2014)', 'Py Torch-style pseudo code', and refers to various frameworks like 'Functa', 'Composers', 'Vi T'. However, it does not provide specific version numbers for any software libraries or frameworks used, such as PyTorch version or Python version. |
| Experiment Setup | Yes | In all experiments, we use Adam optimizer (Kingma & Ba, 2014) for meta-optimization, with a learning rate of 10 4 and the momentum parameters are set as (β1, β2) = (0.9, 0.999). ... Table 4: List of common configurations for each dataset. ... batch size 64 32 16 32, epoch 16,000 300 300 300, total inner step K for optimization-based methods 3, width/depth of INRs 128/5, σ for fourier feature None 30.0 30.0 30.0. |