Inference process

[hehulingxiao]

Thanks for your great work!

I have a question: How can we use the model for classification during inference? Should we use image embeddings after cross-attn module or just from the image encoder? Do we need to compare the embeddings in the hyperbolic space, or just compare them in the original Euclidean space?

Best regards

[WeiWu2025]

Thanks for your interest in our work and for the great question！

Yes, during inference we use the image embeddings after the cross-attention module, consistent with the alignment stage in our method. The cross-attention module is designed to construct hierarchy-specific visual features that better match the hierarchical textual features.

For the geometry part, our current backbone features are still produced in the Euclidean space. We treat them as features in the tangent space, and project them to the hyperbolic space mainly for the modality-alignment constraints during training. At inference time, we compare features in the tangent space, rather than directly using a hyperbolic metric.

Exploring a more direct use of hyperbolic geometry during inference could also be an interesting direction, and we may study this in future work.

Best regards