Interactive demo for explaining CNN image classification using Grad-CAM (Gradient-weighted Class Activation Mapping).
This project allows users to:
- Upload an image
- Classify it using a ResNet-18 convolutional neural network
- View the top-3 predicted ImageNet classes
- Visualize a Grad-CAM heatmap highlighting which regions influenced the model’s prediction
The demo is deployed as an interactive web application using Gradio and hosted on Hugging Face Spaces.
Modern deep learning models often behave like black boxes, making it difficult to understand why a particular prediction was made.
The goal of this project is to explore model interpretability for convolutional neural networks by visualizing which image regions influence predictions.
Using Grad-CAM, we can trace how gradients flow through convolutional feature maps and produce a heatmap highlighting the parts of the image most responsible for the model's decision.
This helps verify whether the model is focusing on meaningful visual features or spurious background patterns.
Try the interactive application here:
https://huggingface.co/spaces/dschechter27/vision_model_interpretability
Upload an image and instantly see the model’s prediction and visual explanation.
Model prediction: liner
Grad-CAM highlights the structure of the ship, showing that the model focuses on the correct object region.
Model prediction: Bernese Mountain Dog
The heatmap concentrates on the dog’s body and head, demonstrating that the model uses meaningful visual features when making its prediction.
Model prediction: sports car
Grad-CAM focuses on the car body and wheels, confirming that the model is identifying relevant vehicle features.
Grad-CAM visualizations suggest that the CNN relies on localized semantic features when making predictions.
Across several examples:
• For ships, the heatmap concentrates on the hull and upper deck structure
• For dogs, the model focuses on the head and torso
• For cars, the network highlights the body and wheel regions
These results indicate that the network is learning meaningful visual features rather than relying primarily on background artifacts.
However, Grad-CAM can sometimes highlight broad regions rather than precise object boundaries, which reflects the spatial resolution limitations of convolutional feature maps.
Grad-CAM explains CNN predictions by analyzing the gradient of the predicted class score with respect to convolutional feature maps.
flowchart LR
A[Input Image] --> B[ResNet-18 CNN]
B --> C[Feature Maps A_k]
B --> D[Class Score y]
D --> E[Backpropagation]
E --> F[Gradients ∂y/∂A_k]
F --> G["Compute weights
α_k = average( ∂y / ∂A_k )"]
G --> H["Weighted feature maps
Σ_k α_k · A_k"]
H --> I["Apply ReLU
L_gradcam = ReLU( Σ_k α_k · A_k )"]
I --> J[Grad-CAM Heatmap]
J --> K[Overlay Heatmap on Image]
K --> L[Visual Explanation]
flowchart LR
A[Upload Image] --> B[Image Preprocessing]
B --> C[ResNet-18 CNN]
C --> D[Class Prediction]
D --> E[Compute Gradients]
E --> F[Grad-CAM]
F --> G[Heatmap Generation]
G --> H[Overlay Heatmap on Image]
H --> I[Visual Explanation]
Image
→ ResNet-18 Prediction
→ Gradient Backpropagation
→ Grad-CAM Heatmap
→ Visual Explanation
vision-model-interpretability/
assets/ # README example screenshots
images/ # original input images
results/ # Grad-CAM outputs
app.py # Gradio web application
requirements.txt # project dependencies
vision_model_interpretability.ipynb # development notebook
README.md
- PyTorch
- Torchvision
- Grad-CAM
- Gradio
- OpenCV
- NumPy
- Hugging Face Spaces
Install dependencies:
- pip install -r requirements.txt
Run the web app:
- python app.py
Then open the Gradio interface in your browser to upload images and visualize model explanations.
David Schechter
MIT Class of 2030
Interested in machine learning, AI interpretability, and computer vision.
GitHub: https://github.com/dschechter27875
MIT License