• Mehdi:

Also validating for robustness, generalization is key, especially that many works exist that solve similar problems but are SUPERVISED (for example the old work of Yusuke).

Here we are using SSL-JEPA who helps to generalize, building world models etc, and our work should be the first work that:

• solves distributed crossmodal prediction
• solves distributed crossmodal ISAC;
• role of RF/CSI. use CSI embedding to predict image embedding (using actual CSI may belimiting. unless you want to use for conditioninig for the JEPA image prediction part) for the actual validation, hopefully we can show various tasks (at least two)..if we only show classification it is classical and limiting.. we should validate Visual Q&A  and/or RF Q&A using crossmodality trick—-> this has never been done

• Nishio:

さて、その位置付けを強調するには、

1. 自己教師あり学習にはunlabeled データが大量に利用可能

2. 教師あり学習によるdown stream tasksには、少量のラベルしか利用できないという設定が不可欠になるでしょう

3. 自己教師あり学習にはunlabeled データが大量に利用可能

4. 教師あり学習によるdown stream tasksには、少量のラベルしか利用できない