This project was developed as part of my Master's program in Data Science and Artificial Intelligence. It focuses on classifying images from the CIFAR-10 dataset using Convolutional Neural Networks (CNNs). The primary goal was to achieve at least 80% accuracy on the test set through architectural modifications alone. I conducted a total of 12 experiments, exceeding the minimum requirement of 10 different experiments.
The CIFAR-10 dataset consists of 60,000 32x32 color images across 10 classes. The goal of this project was to enhance the performance of a very basic CNN model provided by the professors, implementing 12 different experiments (in addition to the baseline model) to improve its accuracy.
- Baseline Accuracy: 60.6% on the test set.
- Best Accuracy without Data Augmentation: 87.7%.
- Best Overall Accuracy: 90.7% using Data Augmentation.
- Transfer Learning: The final experiment applied VGG-16, achieving lower accuracy than the augmented CNN. Future improvements will focus on better applying transfer learning.
- Deep Learning: Implementation and optimization of CNN architectures.
- Model Optimization: Techniques like dropout, batch normalization, and data augmentation.
- Data Augmentation: Applied to improve model generalization and performance.
- Transfer Learning: Initial exploration using VGG-16, with plans to improve this aspect in future work.
- Keras: All models and experiments were implemented using the Keras library.
/notebooks: Jupyter notebooks for each experiment conducted, detailing modifications and results./docs: Final project report (PDF) with detailed explanations of each experiment, results, and future developments.
The CIFAR-10 dataset is available through TensorFlow Datasets or can be downloaded from the official page.
The project successfully reached an accuracy of 90.7% using Data Augmentation. However, there is still room for improvement, especially in the area of Transfer Learning. The VGG-16 model used in the final experiment did not surpass the results obtained with Data Augmentation, and further study in this area could yield better results.
Feel free to reach out for any questions or further discussions!
This project is licensed under the MIT License - see the LICENSE file for details.