Multimodal Classification via Total Correlation Maximization
Feng Yu Xiangyu Wu Yang Yang Jianfeng Lu Nanjing University of Science and Technology · 2026
Read Multimodal Classification via Total Correlation Maximization by Feng Yu Xiangyu Wu Yang Yang Jianfeng Lu Nanjing University of Science and Technology (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-13015.txt.
Description
Many-mode learning should combine sensors, but joint training often overfits strong modalities and neglects weak ones, sometimes losing to single-modality models. Feng Yu and three co-authors study modality competition information-theoretically and propose maximising total link between many-mode traits and labels. That objective eases competition while capturing interactions between modalities through aligned traits.
Drawing on mutual-information estimation, they present total link neural estimation to bound total link from below. The final loss, TCMax, is hyperparameter-free and optimises the varying bound straight. Wide tests show TCMax beating state-of-the-art joint and unimodal approaches. For many-mode sorting, the paper offers both a diagnosis of competition and a principled, tuning-free remedy.
Catalog record
| CREATOR | Feng Yu Xiangyu Wu Yang Yang Jianfeng Lu Nanjing University of Science and Technology |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-13015