Interpretable Visual Semantic Alignment via Spectral Attribution

Our work specifically focuses on generating visual explanations of both the image modality and the text modality using concepts from graph spectral theory by solving the characteristic equation of the Laplacian for models that are fine-tuned on the VQA downstream task. Our method is model-agnostic to some extent and post-hoc, capable of generating local explanations catering to each image-question pair provided for inference. The job of our interpretability module is to highlight the regions of image and text (question) that contribute to the model’s decision/answer. We propose the following list of contributions for interpretability of bi-modal models:

1. Class-agnostic version (Deep Spectral Methods for Interpretability - DSMI): The first is the direct application of graph spectral theory which is unsupervised in nature

2. Class-specific version (DSMI + Grad): In order to explain the decision of the bi-modal models, we modeled the spectral approach like GradCAM, where "CAM" represents the fiedler eigenvectors generated by the standard spectral approach which are much more robust than pure attention maps

3. Class-specific and Transformers-specific version (DSMI + Grad + Attn): In combination with the gradients based spectral approach, we utilise the attention maps for transformer specificit

Demo

git clone https://github.com/shiv2110/METER-spectral-interpretability.git
cd METER-spectral-interpretability
pip install -r requirements.txt

Download Pre-trained Checkpoint

Here are the pre-trained models: METER-CLIP16-RoBERTa fine-tuned on VQAv2 (resolution: 576^2) meter_clip16_288_roberta_vqa.ckpt

 python demo_vqa_ours.py with num_gpus=0 load_path=<path_to_ckpt_file> test_only=True method_name=<interpretability_method> img=<image_path> question=<question>

Visualisation of explanations provided by DSMI variants on LXMERT

Running Perturbation Tests

Download VQA 2.0 validation dataset

https://visualqa.org/download.html

Download COCO 2014 val images, , annotations (val), and questions (val)

root
├── val2014              
|   ├── COCO_val2014_000000000042.jpg
|   └── ...  
├── v2_OpenEnded_mscoco_val2014_questions.json
└── v2_mscoco_val2014_annotations.json

Download valid.json

cd data/vqa
wget https://nlp.cs.unc.edu/data/lxmert_data/vqa/valid.json

Running positive image perturbation test

python perturbation.py with COCO_path=<COCO_path> method_name=<method_name> is_positive_pert=True modality="image" test_type="positive"

Running negative image perturbation test

python perturbation.py with COCO_path=<COCO_path> method_name=<method_name> is_positive_pert=False modality="image" test_type="negative"

Running positive text perturbation test

python perturbation.py with COCO_path=<COCO_path> method_name=<method_name> is_positive_pert=True modality="text" test_type="positive"

Running negative text perturbation test

python perturbation.py with COCO_path=<COCO_path> method_name=<method_name> is_positive_pert=False modality="text" test_type="negative"

AUC scores for perturbations tests on METER: Lower AUC for positive tests and higher AUC for negative tests indicates better performance

Method	Image ve+	Image ve-	Text ve+	Text ve-
Hila Chefer Relevance Maps	55.27	82.97	31.70	61.33
Transformer Attribution w/o LRP	55.26	82.96	31.72	61.29
Raw Attention	55.33	82.49	36.72	54.41
Grad-CAM	70.95	73.52	49.21	49.83
Rollout	59.96	82.17	44.32	49.93
DSMI	60.07	79.47	45.14	48.05
DSMI + Grad	54.45	82.93	34.75	57.64
DSMI + Grad + Attn	55.45	82.45	40.37	52.38

Credits

• Implementation of baselines is based on Relevance Maps
• Core DSMI implementation is based on Deep Spectral Methods for unsupervised localisation and segmentation
• Implementation of METER is based on METER