Interpreting CLIP with Hierarchical Sparse Autoencoders
Paper
• 2502.20578 • Published
MSAE FashionCLIP (4096)
This is a Matryoshka Sparse Autoencoder (MSAE) trained on CLIP image embeddings from the KAGL dataset.
Model details
- Input dim: 512 (CLIP ViT-B/32 image embeddings)
- Hidden dim: 4096
- Sparsity: Matryoshka
- Training dataset size: 39,990 images
- Training objective: reconstruction + sparsity
- Two checkpoints, one at 30 epochs and another at 100.
Usage
This model is intended for interpretability and feature analysis of CLIP embeddings.
Citation
The matryoshka SAE architecture comes from the following paper: https://arxiv.org/abs/2502.20578
Code from this paper's repository was used for the training.
- Downloads last month
- 6
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for Alloryne/msae-clip-fashion-4096-topk64
Base model
patrickjohncyh/fashion-clip