This project involved a comprehensive analysis comparing Support Vector Machine (SVM) kernels against a Feedforward Neural Network (MLP) baseline for multi-class image classification on the CIFAR-10 dataset. The implementation included two distinct SVM methods: a non-linear RBF Kernel requiring Principal Component Analysis (PCA) for dimensionality reduction (achieving 43.91% Test Accuracy), and a Linear Kernel baseline (achieving 35.48% Test Accuracy). All models were implemented using Scikit-Learn and TensorFlow for data handling. The extensive experimentation documented the significant performance gap between kernels, highlighting the inability of linear models to capture complex image features and the superiority of non-linear mapping for visual data.

View on GitHub

This project involved a comprehensive analysis comparing different Feedforward Neural Network (FNN) architectures for multi-class image classification on the CIFAR-10 dataset. The implementation included two distinct methods: an MLP Baseline requiring Principal Component Analysis (PCA) for dimensionality reduction (achieving 52.88% Test Accuracy), and a Convolutional Neural Network (CNN) (achieving 70.00% Test Accuracy). All models were trained using the Back-propagation algorithm with the Adam optimizer. The extensive experimentation documented Overfitting in both architectures, highlighting the critical need for regularization and Early Stopping in production models

View on GitHub

Built a compact, regularized TensorFlow neural network to forecast NO2, CO, SO2, 03, PM2-5, PM10 particles over Greece and deliver predictions to an interactive 3D globe. Data was taken from NASA's GIOVANNI interface. Trained on a standardized (z-scored) target, and used gradient clipping to keep learning stable. Generalization was enforced with out-of-year validation (train on earlier years, test on the last year), while ReduceLROnPlateau and EarlyStopping (best-weights restore) prevented overfitting and landed training at the strongest epoch. Project was developed over 2 days for NASA's Space Apps Challenge Hackathon in Thessaloniki where i worked together with 5 other people and integrated the neural network on an interactive webpage.

View on GitHub