Основные понятия
The author reveals the subject variation problem in multimodal human intention understanding and introduces a Subject Causal Intervention module (SuCI) to address prediction bias caused by subject-specific spurious correlations.
Аннотация
This study delves into the subject variation problem in multimodal human intention understanding, highlighting the detrimental impact of subject-specific spurious correlations on model predictions. The proposed SuCI module aims to mitigate this bias by disentangling subject confounders and achieving unbiased predictions across different subjects. Extensive experiments on various benchmarks demonstrate the effectiveness of SuCI in improving model performance and generalizability.
Key Points:
- Multimodal intention understanding faces challenges due to subject variation.
- SuCI is introduced to remove prediction bias caused by subject-specific correlations.
- Experiments show significant improvements with SuCI across different benchmarks.
Статистика
Existing MIU models suffer from a prediction bias due to subject variation.
SuCI achieves new SOTA performance on multiple benchmarks.
Цитаты
"As a research hotspot that combines linguistic and non-verbal behaviors, multimodal intention understanding has attracted significant attention."
"Motivated by this observation, we introduce a recapitulative causal graph to formulate the MIU procedure and analyze the confounding effect of subjects."