bonus point (the feature extractor is available on HF): huggingface.co/DBD-research...

[10/10] Wrap-up 🎯
🔹 Unified supervised + unsupervised hashing
🔹 Flexible: works via probing or LoRA
🔹 SOTA hashing in minutes on a single GPU

📄 Paper: arxiv.org/abs/2510.27584
💻 Code: github.com/ilyassmoumma...

Shoutout to my wonderful co-authors Kawtar, Hervé, and Alexis.

[9/10] Strong generalization 🌍
CroVCA produces compact codes that transfer efficiently:
✅ Single HashCoder trained on ImageNet-1k works on downstream datasets without retraining (More experiments and ablations in the paper)

[8/10] Semantically consistent retrieval 🔍
CroVCA retrieves correct classes even for fine-grained or ambiguous queries (e.g., indigo bird, grey langur).
✅ Outperforms Hashing-Baseline
✅ Works with only 16 bits and without supervision

[7/10] Compact yet meaningful codes 💾
Even with just 16 bits, CroVCA preserves class structure.
t-SNE on CIFAR-10 shows clear, separable clusters — almost identical to the original 768-dim embeddings.

[6/10] Strong performance across encoders 💪
Tested on multiple vision encoders (SimDINOv2, DINOv2, DFN…), CroVCA achieves SOTA unsupervised hashing:

[5/10] Fast convergence 🚀
CroVCA trains in just ~5 epochs:
✅ COCO (unsupervised) <2 min
✅ ImageNet100 (supervised) ~3 min
✅ Single GPU
Despite simplicity, it achieves state-of-the-art retrieval performance.

[4/10] HashCoder 🛠️
A lightweight MLP with final BatchNorm for balanced bits (inspired by OrthoHash). Can be used as:
🔹 Probe on frozen features
🔹 LoRA-based fine-tuning for efficient encoder adaptation

[3/10] Unifying hashing 🔄
Can supervised + unsupervised hashing be done in one framework?

CroVCA aligns binary codes across semantically consistent views:
Augmentations → unsupervised
Class-consistent samples → supervised

🧩 One BCE loss + coding-rate regularizer

[2/10] The challenge ⚡
Foundation models (DINOv3, DFN, SWAG…) produce rich embeddings, but similarity search in high-dimensional spaces is expensive.
Hashing provides fast Hamming-distance search, yet most deep hashing methods are complex, slow, and tied to a single paradigm.

I heard that the Linux client is buggy, I use it on the browser and it's working ok.

for the curious, the code, slides and the article are on Github: github.com/BastienPasde...

love it haha wish I were there to hear Prostitute Disfigurement in an amphitheater

Im interested in the quantum and footnotesize, how much params should they have 😂

It feels like we can now fit more noise with more model capacity 🤔 (Figure 6), maybe we need newer architectures and/or newer training losses.

👋 I worked on bioacoustics during my PhD, but I post mostly about AI

Congratz! 👏

my new addiction today: youtu.be/dSyJqwN36ow
I can't wait to see them this summer in Motocultor Festival