SOTA SSL methods like SwAV, DINO, and VICRegL utilize more than two views of the same data point. But they handle them suboptimally - just aggregating pairwise losses!
This creates conflicting objectives & misses crucial interactions.
We fix this π§΅π
https://t.co/n76ZtoIons
Happy to share our work "M3: Multimodal Masking applied to Sentiment Analysis" which we presented at Interspeech 2021, Brno.
Paper: https://t.co/p6htP5Xa7c
Code: https://t.co/fxiITWln1n
Getting ready for @interspeech2019 where we'll be presenting our paper "Deep Hierarchical Fusion with application in Sentiment Analysis" in a poster session.