RobustMNIST v1.0

--- license: apache-2.0 pipeline_tag: image-classification library_name: pytorch datasets: - mnist tags: - mnist - robust - open-set - computer-vision - safetensors model-index: - name: RobustMNIST-v1.0 results: - task: type: image-classification dataset: name: MNIST type: mnist metrics: - name: Accuracy (Clean) type: accuracy value: 99.51 - name: Accuracy (Extreme OOD) type: accuracy value: 92.33 --- # RobustMNIST (v1.0) <div align="left"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64b433c3faa3181a5e98c87c/nex9yBr2wmq88q9UwHULO.png" width="1000"> </div> `RobustMNIST` is a lightweight, 11-class convolutional neural network designed for handwritten digit recognition. Unlike standard models, this architecture is built to handle out-of-distribution (OOD) inputs and extreme image corruption through a dedicated **"Unknown"** class. This repository contains the model weights serialized in the secure **Safetensors** format (`model.safetensors`), eliminating the security risks associated with standard PyTorch pickle imports. ### Model Details * **Developed by:** MultivexAI * **Version:** 1.0 * **Task:** Open-Set Handwritten Digit Recognition * **Architecture:** 6-Layer Gated CNN (approx. 430k parameters) * **Classes:** 11 (0–9 for standard digits, **10 for "Unknown"**) * **Input:** 1x28x28 grayscale image. --- ## The "Unknown" Class (Class 10) Traditional MNIST models often guess a digit confidently even when the input is just random noise or a shape that isn't a number. RobustMNIST introduces **Class 10**, representing the "Unknown" domain. - **In-Distribution:** For clean digits, the model predicts classes 0–9 with high accuracy, while maintaining a 15-20% uncertainty margin for Class 10. - **Out-of-Distribution:** When an image is severely corrupted (noise, stains, blurs) or represents a non-digit shape, the model's confidence shifts entirely to Class 10. --- ## Performance Metrics Evaluation on standard MNIST and extreme corruption sets: | Set | Accuracy | | :--- | :--- | | **Clean MNIST Test Set** | 99.51% | | **Extreme OOD / Corrupted Set** | 92.33% | --- ### Limitations & Expectations While titled **RobustMNIST**, it is important to clarify that "robust" does not mean "invincible." This is a small-scale model designed to demonstrate OOD detection, not a perfect safety system. - **No 100% Guarantee:** Like...