SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability

📊 Slides

SPARC (Sparse Autoencoders for Aligned Representation of Concepts) learns a unified latent space shared across diverse models (DINO, CLIP Vision, CLIP Text) where individual dimensions correspond to the same concepts across different architectures and modalities. This enables cross-model interpretability, allowing you to understand what concepts are being captured consistently across different vision and vision-language models.

Installation

git clone https://github.com/AtlasAnalyticsLab/SPARC.git
cd SPARC
pip install -e .

Architecture

The SPARC architecture processes multiple input streams (CLIP Text, CLIP Image, DINO Image) through stream-specific encoders and decoders, coupled by a shared sparse activation mechanism that promotes concept alignment.

Key Components:

• Stream Encoders: Transform input features x^s into pre-activation logits h^s through affine transformations
• Global TopK: Aggregates logits across all streams, selects top-k indices globally, then applies the same indices to create sparse representations z^s for each stream
• Stream Decoders: Reconstruct original features from sparse latents using stream-specific linear transformations

The Global TopK mechanism ensures identical TopK selection across all streams for each sample, enforcing structural sparsity and cross-modal alignment. However, this only constrains which latents are selected, not their activations or semantic content. Cross-reconstruction loss is therefore essential to ensure individual latents encode similar concepts across modalities.

Feature Extraction

Option 1: Download Pre-computed Features

Download pre-extracted features and trained models:

Download Features & Results

Contains:

• features/ - Pre-extracted DINO and CLIP features for Open Images and COCO
• final_results/ - Pre-trained SPARC model weights and analysis results

Option 2: Extract Your Own Features

COCO Dataset

# Extract both DINO and CLIP features for COCO validation set
python sparc/feature_extract/extract_coco.py \
    --split val \
    --data_root ./dataset/COCO \
    --output_dir_base ./features/coco \
    --extract_dino --extract_clip

# Extract only DINO features with custom model
python sparc/feature_extract/extract_coco.py \
    --split train \
    --dino_model dinov2_vitl14_reg \
    --batch_size 64 \
    --extract_dino

Open Images Dataset

# Extract features for training split  
python sparc/feature_extract/extract_open_images.py --model clip --split train

# Extract features for test split
python sparc/feature_extract/extract_open_images.py --model dinov2 --split test

Do the same for other splits and models.

Running SPARC Training and Evaluation

Training SPARC Models

# Train SPARC model with global TopK on COCO
python main.py \
    --config configs/config_coco.json \
    --topk_type global \
    --n_latents 8192 \
    --k 64 \
    --cross_loss_coef 1.0

# Train with local TopK (per-modality)
python main.py \
    --config configs/config_coco.json \
    --topk_type local \
    --n_latents 16384 \
    --k 128

Evaluation Only

# Run evaluation with pre-trained COCO model (runs both classification and retrieval)
python main.py --config configs/config_coco.json --only_eval

# Run evaluation with pre-trained Open Images model  
python main.py --config configs/config_open_images.json --only_eval

Key Parameters

• --topk_type: global (shared concepts) or local (per-modality concepts)
• --n_latents: Number of latent dimensions (typically 8192-16384)
• --k: Number of active latents per sample (sparsity level)
• --cross_loss_coef: Weight for cross-modal alignment loss

Notebooks and Analysis

To ensure that the notebooks are running, download the model weights from Download Features & Results.

Qualitative Analysis

notebooks/qual/t1_visualize_dim.ipynb

• Use this notebook to visualize individual latent dimensions across different configurations (Global TopK vs Local).
• See example results in VISUALIZATIONS

notebooks/qual/t2_aligned_concept_visualization.ipynb

• This notebook generates heatmaps/attribution scores for both image and text modalities based on individual or a set of concept-specific latents.
• See example results in VISUALIZATIONS

notebooks/qual/t3_heatmaps_from_caption.ipynb

• This notebook generates heatmaps/attribution scores for both image and text modalities based on similarity of shared latents for each modality to the other.
• See example results in VISUALIZATIONS

notebooks/qual/scripts_all_results/t1_latent_visualizations_open.py

• Latent visualization script specifically for Open Images dataset
• Batch processing for generating latent activation visualizations

notebooks/qual/scripts_all_results/t2_concept_batch_visualization.py

• Batch processing script for concept visualization across multiple images

notebooks/qual/scripts_all_results/t3_caption_visualization.py

• Batch processing multiple text inputs to generate corresponding visual attributions

notebooks/qual/t456_retrieval.ipynb

• Retrieval analysis and evaluation using SPARC latents for in-distribution tasks (meaning each sample is guaranteed to have a corresponding pair) on test set of Open Images

notebooks/qual/t789_retrieval_ood.ipynb

• Out-of-distribution retrieval analysis

Quantitative Analysis

notebooks/quant/t1_jaccard_dead_latents.ipynb

• Jaccard similarity analysis between latent activations
• Dead latent calculation and analysis across different configurations

notebooks/quant/t2_probe_1d.ipynb

• 1D linear probing analysis on SPARC latents
• Evaluates interpretability and linear separability of learned concepts

notebooks/quant/t3_segmentation-coco-latents-global.ipynb

• Semantic segmentation evaluation using Global TopK SPARC latents

notebooks/quant/t3_segmentation-coco-latents-local.ipynb

• Semantic segmentation evaluation using Local TopK SPARC latents

notebooks/quant/t3_segmentation-coco-caption.ipynb

• Caption-based segmentation evaluation
• Tests cross-modal concept transfer from text to visual segmentation
• Validates that text-derived concepts can localize visual regions

notebooks/quant/t3_segmentation-coco-DETR.ipynb

• DETR-based segmentation as a baseline for evaluation

notebooks/quant/t4_retrieval_clip_baseline.ipynb

• CLIP baseline comparison for retrieval tasks

For comprehensive visualization examples and results from all notebooks, see VISUALIZATIONS.

Citation

@misc{nasirisarvi2025sparcconceptalignedsparseautoencoders,
      title={SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability}, 
      author={Ali Nasiri-Sarvi and Hassan Rivaz and Mahdi S. Hosseini},
      year={2025},
      eprint={2507.06265},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2507.06265}, 
}

Contact

• GitHub: AtlasAnalyticsLab/SPARC
• Email: ali.nasirisarvi@mail.concordia.ca