CNN-Based Image Classification and Uncertainty Analysis Using CIFAR-10 Dataset

Module Information

Module: 6009CMD – Artificial Neural Networks (Coventry University)
Assessment: Individual Coursework (Artificial Neural Networks Application)
Author: Xingkai Feng
Student ID: 16038226
Video OneDrive link (view-only): https://livecoventryac-my.sharepoint.com/:v:/r/personal/fengx29_uni_coventry_ac_uk/Documents/Artificial%20Neural%20Network_XingkaiFeng_16038226/video/6009CMD_XingkaiFeng_Video.mp4?csf=1&web=1&nav=eyJyZWZlcnJhbEluZm8iOnsicmVmZXJyYWxBcHAiOiJPbmVEcml2ZUZvckJ1c2luZXNzIiwicmVmZXJyYWxBcHBQbGF0Zm9ybSI6IldlYiIsInJlZmVycmFsTW9kZSI6InZpZXciLCJyZWZlcnJhbFZpZXciOiJNeUZpbGVzTGlua0NvcHkifX0&e=KwLivz

Project Overview

This project compares two Convolutional Neural Network (CNN) approaches for CIFAR-10 image classification and includes an uncertainty/ambiguity analysis to highlight difficult-to-classify samples.

Part A (Human-guided): A baseline CNN designed and implemented using my own understanding and research.
Part B (GenAI-assisted): An improved CNN architecture developed with GenAI guidance (e.g., adding Batch Normalisation and Dropout), then reviewed, implemented, and evaluated by me.
Uncertainty Analysis: Low-confidence predictions are flagged as ambiguous to indicate uncertain or potentially misclassified inputs.

Dataset

Dataset: CIFAR-10 (60,000 colour images, 10 classes)
Split: 50,000 training images / 10,000 test images
Access method: Loaded via TensorFlow/Keras API (cifar10.load_data()), cached locally

Final Results (Test Set)

Part A (Baseline CNN)

Accuracy: 0.6980
Loss: 1.1201
Ambiguous samples (low-confidence): 1572

Part B (Improved CNN)

Accuracy: 0.7884
Loss: 0.6340
Ambiguous samples (low-confidence): 2086

Note: Although Part B achieves higher accuracy, it flags more ambiguous samples under the same confidence threshold, suggesting a more cautious confidence distribution on difficult inputs.

How to Run

1) Install dependencies

If you have requirements.txt:

pip install -r requirements.txt

fengx29/6009CMD_ANN_CIFAR10_Xingkai_Feng

About

Resources

Stars

Watchers

Forks

Releases

Languages