Goal: To explore the potential of evolutionary search methods in designing robust neural network architectures.
This project focuses on Neural Architecture Search (NAS) using Genetic Algorithms (GA) to optimise Convolutional Neural Networks (CNNs) for classifying Farsi vowel audio signals. The goal was to discover high-performing architectures without manual tuning.
Approach:
The GA evolves a population of CNN architectures through selection, crossover, and mutation. Each individual is trained and evaluated based on validation accuracy.
Dataset:
Farsi vowel audio dataset with 6 classes, preprocessed from MATLAB .mat files into PyTorch-ready tensors.
Results:
Achieved a validation accuracy significantly above random guessing, showing the GA’s effectiveness in architecture discovery. Hyperparameter tuning (e.g., mutation rate) revealed its impact on convergence and diversity.
Tech stack: Python, PyTorch, NumPy, matplotlib