[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:

[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

They get rid off many tricks that help prevent collapse in DINO. Hoping that this will pave the way for more simplified solutions derived from principled approaches!

By applying MaskCut on models trained with both DINO and SimDINO, they confirm similar emerging properties.

They show that SimDINO is more stable during training as k-NN accuracy on ImageNet keeps improving. Additionally, when training on a different dataset (COCO), SimDINO is more robust.

SimDINO and SimDINOv2 outperform DINO and DINOv2 both in classification and segmentation tasks. Additional experiments are provided in the paper.

The regularization term measures the size of the covariance matrix of the embedding examples. It was proposed in 2007 by Ma to segment multivariate mixed data by minimizing the coding length (subject to distortion). Many great ideas and proofs can be found here: ieeexplore.ieee.org/document/428...

Yi Ma & colleagues managed to simplify DINO & DINOv2 by removing many ingredients and adding a robust regularization term from information theory (coding rate) that learn informative decorrelated features. Happy to see principled approaches advance deep representation learning!

A really cool paper from Kyutai demonstrates how model capabilities can be extended to a new domain (e.g., learning a new language) while preserving the original capabilities. This is achieved by leveraging the concept of adapters.