In this session at the Data Innovation Summit 2024, we have Binwei Yang from Walmart Global Tech! In his talk, Binwei addresses the fundamental challenge of fine-grained attribute classification for unseen objects, proposing a novel method for learning disentangled representations. By breaking down prompts into sub-prompts and performing a unique late fusion of output embeddings, we achieve composable attribute representations that transfer to both seen and unseen objects during inference. Key takeaways:
- Leveraging natural language supervision, we introduce an effective strategy for learning generalizable, disentangled attribute representations.
- By formulating the fine-grained attribute classification task, we assess the transferability and composability of learned representations for unseen objects.
- Our continuous prompting method, tested extensively on diverse datasets in a few-shot setting, demonstrates significant improvements over existing methods.
Binwei’s presentation, comments and opinions are provided in their personal capacity and not as a representative of Walmart. They do not reflect the views of Walmart and are not endorsed by Walmart.
