Welcome to Team Synapse UoM's GitHub organization! We are a group of electrical engineering students from the University of Moratuwa, working on the Pattern Recognition (EN3150) Group Assignment. Our focus is on deep learning techniques - Convolutional Neural Networks (CNNs) for image classification.

Repository: Custom-CNN-ImageClassification

In this project, all four team members collaborated to design and implement a custom CNN model for the classification of handwritten digits from the MNIST dataset. We followed a structured process to build the model, train it, and evaluate its performance.

• Custom CNN architecture built using Pytorch
• Dataset split into training, validation, and testing sets.
• Parameters tuned for optimal performance.
• Evaluation through accuracy, confusion matrix, precision, and recall.
• Comparisons made with popular architecture.

Each team member has taken the lead in implementing and fine-tuning a specific pre-trained CNN architecture, comparing its performance with the custom model.

• DenseNet Implementation by Gayanukaa
  • DenseNet fine-tuned on our dataset to achieve improved accuracy and model convergence.
• ResNet Implementation by Sulith123
  • ResNet architecture used to handle deeper layers with residual connections.
• GoogLeNet Implementation by nimeshlal097
  • Fine-tuning of GoogLeNet for complex image classification.
• VGG Implementation by SandaruDias
  • VGG architecture optimized to extract more features from the dataset.

Each repository includes:

• Code for loading and fine-tuning the pre-trained models.
• Comparisons between fine-tuned models and the custom CNN.

Phase 1: Initial Setup
Phase 2: Custom CNN Development
Phase 3: Pre-trained Model Fine-tuning
Phase 4: Model Comparison and Report
Phase 5: Final Presentation & Documentation

In addition to the repositories, we are working on a website that will host our project documentation and results.

• Team-Synapse-UoM.github.io (under development)