Note: This project was conducted as part of the Deep LEarning course taught by Professor Kefaya Qaddoum at Concordia University. For further details, insights, and full implementation, please refer to the GitHub repository. The entire workflow and explanations are documented in the report 40316218_DL_Project_Report.
Abstract
This project explores the challenges of training deep learning models in data-scarce scenarios. It is structured in two main parts.
In the first challenge, the task was to classify CIFAR-10 images using only 25 training samples per class, without relying on any external data or pre-trained models. This required carefully tuning small models from scratch, pushing the limits of generalization under severe data constraints.
The second challenge introduced the possibility of leveraging pre-trained networks. I experimented with transfer learning by fine-tuning architectures such as ResNet-18 and ResNet-50. To push further, I also implemented Model-Agnostic Meta-Learning (MAML), enabling the model to quickly adapt to new tasks from limited data.
All experiments were tracked and analyzed using Weights & Biases. This project allowed me to compare different learning strategies—from naive training to meta-learning—while understanding their trade-offs in few-shot learning contexts.