Contrasting Explanation: Comparative Explainability in Deep Learning

📋 Project Overview

This project investigates whether contrastive explanations provide better interpretability than traditional explanations in deep learning models. Rather than answering "Why is this classified as class A?", we explore "Why is this class A rather than class B?".

Research Question

Can providing contrastive information (comparing the predicted class against the closest competing class) lead to more meaningful and interpretable explanations of neural network predictions?

🎯 Motivation

Traditional explainability methods focus on explaining individual predictions. However, human explanations are often contrastive in nature—we naturally explain decisions by comparing alternatives. This project explores whether machine-generated explanations benefit from this contrastive approach.

Key Hypothesis: Comparing feature importance between the top two predicted classes reveals more interpretable patterns than analyzing a single class in isolation.

🔬 Methodology

Explainability Methods Tested

The project implements and compares multiple state-of-the-art explainability techniques:

• Integrated Gradients - Accumulates gradients along a straight line from a baseline to the input
• Noise Tunnel - Computes explanations with smoothing via noise addition
• Gradient SHAP - Combines SHAP with gradient information
• Saliency Maps - Computes input gradients to identify important features

Approaches

1. Classical Explainability: Generate explanations for the predicted class independently
2. Contrasting Explainability: Generate explanations by computing differences between:
   • Feature importance matrix of the predicted class
   • Feature importance matrix of the closest competing class

📊 Datasets

MNIST (Handwritten Digits)

• Performance: ✅ Excellent results
• Example: Distinguishing between 4 and 9
• The contrastive approach clearly highlights the discriminative features between these confusable digits
• The difference matrices reveal structural differences effectively

ImageNet (Real-world Images)

• Performance: ⚠️ Mixed results
• Example: Duck classification
• Contrastive explanations provide less clarity with complex natural images
• Suggests that the approach may be more suitable for simpler, more structured datasets

🔍 Key Findings

Aspect	Classical Approach	Contrasting Approach
MNIST (4 vs 9)	Good	Excellent ✓
MNIST General	Good	Good
ImageNet (Ducks)	Moderate	Moderate
Complex Scenes	Moderate	Moderate
Interpretability	Varies	More focused

💡 Insights

• ✅ Strong for simple, structured data: The contrastive approach excels with datasets like MNIST where classes have clear structural differences
• ⚠️ Limited for complex real-world images: Natural images contain too much contextual information; simple feature differences don't capture semantic distinctions
• 🎯 Best use case: Binary or few-class classification with distinct visual patterns
• 📈 Future improvement: May benefit from hierarchical contrastive analysis or semantic feature grouping

🚀 Usage

Installation

pip install -r requirements.txt

Running the Notebook

jupyter notebook Contrasting_Explanation.ipynb

Project Structure

contrasting_explanation/
├── Contrasting_Explanation.ipynb   # Main analysis notebook
├── README.md                         # This file
├── requirements.txt                  # Python dependencies
├── data/
│   ├── MNIST/                        # Handwritten digits dataset
│   │   └── raw/
│   ├── ImageNet/                     # ImageNet classes
│   │   └── imagenet_class_index.json
│   └── test_image/                   # Sample images for testing
└── weights/
    └── mnist_weights.pth             # Pre-trained MNIST model

📚 Literature Review & References

For comprehensive background on contrastive explanations and interpretability in deep learning, see:

• Contrastive Explanations: Doshi-Velez & Kim (2017) - Towards A Rigorous Science of Interpretable Machine Learning
• Integrated Gradients: Sundararajan et al. (2017)
• SHAP Methods: Lundberg & Lee (2017) - A Unified Approach to Interpreting Model Predictions
• Attention & Saliency: Simonyan et al. (2013) - Deep Inside Convolutional Networks

🔗 Related Work

• Captum Library: PyTorch's interpretability library
• Contrastive Learning: Chen et al. (2020) - SimCLR
• Model Agnostic Meta-Learning (MAML): For few-shot understanding

📝 Future Work

• Test contrastive approach on other structured datasets (medical imaging, document classification)
• Implement hierarchical contrasting (explaining against multiple competing classes)
• Add quantitative metrics for explanation quality
• Extend to NLP models for text classification
• Develop interactive visualization tools
• Compare with recent contrastive learning methods

---

Last Updated: May 2026