Research map / Vision-language and multimodal / Cross-modal learning
Alternating Gradient Descent and Mixture-of-Experts for Integrated Multimodal Perception
Hassan Akbari, Dan Kondratyuk, Yin Cui, Rachel Hornung, Huisheng Wang, Hartwig Adam
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| alternate NCE and softmax objectives instead of summing them (ablation) | IMP on I21K · ImageNet1K linear probe | accuracy: 47.6 -> 49.9 | Improved |
| alternate NCE and softmax objectives instead of summing them (ablation) | IMP on I21K · CIFAR-100 linear probe | accuracy: 82.4 -> 84.1 | Improved |
| add softmax objective alternating with NCE (ablation) | IMP on I21K · ImageNet1K linear probe | accuracy: 39.8 -> 49.9 | Improved |
| sum NCE and softmax objectives instead of NCE only (ablation) | IMP on I21K · CIFAR-100 linear probe | accuracy: 84.0 -> 82.4 | Hurt |
| replace dense with 4-expert experts-choose MoE (ablation) | IMP-B · ImageNet zero-shot | accuracy: 59.8 -> 68.1 | Improved |
| experts-choose routing instead of tokens-choose (ablation) | IMP-B single tower MoE · ImageNet zero-shot | accuracy: 62.5 -> 68.1 | Improved |
| two towers instead of one tower with experts-choose (ablation) | IMP-B MoE · ImageNet zero-shot | accuracy: 68.1 -> 68.4 | No clear effect |
| train-time IMP prompts instead of no prompt (ablation) | IMP · ImageNet zero-shot | accuracy: 51.4 -> 56.5 | Improved |
| train-time IMP prompts instead of no prompt (ablation) | IMP · CIFAR-100 zero-shot | accuracy: 67.5 -> 60.5 | Hurt |
| add I21K with NCE+Softmax to CC pretraining (ablation) | IMP-S · ImageNet linear | accuracy: 15.2 -> 49.0 | Improved |
| add softmax to NCE on CC + I21K (ablation) | IMP-S · ImageNet zero-shot | accuracy: 39.1 -> 49.4 | Improved |
| add LAION to CC + I21K (NCE+Softmax) (ablation) | IMP-S · Flickr30k T->I | recall: 38.6 -> 51.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- CoCa — COCA: COllaborative CAusal Regularization for Audio-Visual Question Answering (AAAI 2023) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Multimodal Ensembling for Zero-Shot Image Classification (AAAI 2024 · similarity 0.88)
- Apollo: An Exploration of Video Understanding in Large Multimodal Models (CVPR 2025 · similarity 0.88)
- 4M-21: An Any-to-Any Vision Model for Tens of Tasks and Modalities (NeurIPS 2024 · similarity 0.88)
- Multimodal Representation Learning by Alternating Unimodal Adaptation (CVPR 2024 · similarity 0.88)
- Fusion or Confusion? Multimodal Complexity Is Not All You Need (NeurIPS 2026 · similarity 0.88)
- DenseFusion-1M: Merging Vision Experts for Comprehensive Multimodal Perception (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
