#ICLR2025 was a blast, great to talk with all of you about our works on hyperbolic vision-language models, a better way to do object detection, and brain-aligned image generation! Next destination: the Web Conference in Australia, happening now

Hyperbolic embeddings not only allow for an asymmetric modeling between vision and language, but open the door to models all sorts of compositions. This paper shows how image-box compositions form cool new hierarchies and strong vision-language models!

We find that hyperbolic vision-language models such as MERU: 1. exhibit spatial awareness (CLIP has no such awareness) 2. can better deal with ambiguity 3. are far better OOD discirminators Despite being trained on the same data, hyperbolic models really learn unique patterns!

Vision-language models benefit from hyperbolic embeddings for standard tasks, but did you know that hyperbolic vision-language models also have surprising properties? Our new #TMLR paper shows 3 intriguing properties. w/ @sarahibrahimi_ @GhadimiAtigMina @nannevn @marcelworring

Hyperbolic Random Forests at been accepted to #TMLR! Random forests split data hierarchically, making for strong links with hyperbolic space. We show how to generalize forests through horospheres, resulting in powerful machine learning models. Paper: openreview.net/pdf?id=pjKcIz…

8 "G3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding" by Yuan Wang et al. Hyperbolic learning allows for a better representation space to deal with the complex spatial and semantic relations in 3D visual grounding.

7 "HyperSDFusion: Bridging Hierarchical Structures in Language and Geometry for Enhanced 3D Text2Shape Generation" by Zhiying Leng et al. 3D shape generation is strongly hierarchical. This paper shows how to use such structures in hyperbolic space to get impressive 3D shapes.

6 "Accept the Modality Gap: An Exploration in the Hyperbolic Space" by Sameera Ramasinghe et al. VLMs should operate in hyperbolic space! This paper shows how to preserve intra-modal hierarchies and cross-modal alignment in hyperbolic space.

5. "Rethinking Generalizable Face Anti-spoofing via Hierarchical Prototype-guided Distribution Refinement in Hyperbolic Space" by Chengyang Hu et al. Hierarchical hyperbolic learning is beneficial in face recognition as well, especially with their proposed prototype learner.

4. "Hyperbolic Learning with Synthetic Captions for Open-World Detection" by Fanjie Kong et al. Open world detection is a hot new topic. This paper bootstraps VLMs with synthetic captions to get rich descriptions. Hyperbolic learning is a great tool to deal with noisy captions.

3. "Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping" by @gwonhye67424680 et al. Visual scenes are naturally hierarchical, but how to exploit this structure in deep networks with hyperbolic space? This paper proposes a hyperbolic mapper of image parts.

2. "Hyperbolic Anomaly Detection" by Huimin Li et al. Performing distance-based learning in hyperbolic space boosts anomaly detection, no special tricks needed! Hyperbolic space seems to be naturally well-equiped to deal with robustness, from anomalies to OOD.

1. "Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincaré Ball" by @SimWeberTUM et al. Standard segmentation suffers from a bias to the parent class. Hyperbolic space is not biased due to more uniform separation of class margins, no hierarchies needed!

Our solution is to take a transductive view by matching the distributions of embedded actions and videos. We do so through a hyperspherical optimal transport. The proposed approach is generic and works for any zero-shot method, improving recognition and localization. 3/3

Look at the figure below. The orange line denotes how often each action is selected during testing for a SOTA large video model. This is a clear long-tailed distribution with a few popular actions. As such, the accuracy for most actions is low or zero, hampering accuracy. 2/3

In unsupervised literature, we also focus on three main trend, namely (i) generative learning, (ii) clustering, and (iii) self-supervised learning. 3/4