PRELUDIUM 23
Project: Solve challenges in Data-Centric AI using advanced sampling techniques
Implementation period: 05.02.2025 - 04.02.2027
Project budget: 119 560 PLN
Deep learning has achieved significant successes in various fields, mainly due to the availability of large data sets and increased computing power. However, the quality of training data remains a key factor affecting the performance of models. For this reason, the Data-Centric AI approach, which emphasizes high data quality and efficient training processes, is becoming increasingly important.
The goal of our project is to use data sample selection methods to solve important problems in Data-Centric AI, such as robust training of models on noisy labeling problems, training resistant to adversarial attacks, and explaining the learning process. These problems have not yet been thoroughly studied in the literature, which makes our project pioneering.
We undertook this research topic because optimizing data usage is crucial due to the large computational costs associated with training deep learning models. Selecting the right data samples for training can significantly increase the robustness of models and optimize resource utilization. Methods such as Importance Sampling, Selective Backprop or streaming approaches have already shown that effective sample selection can improve the performance and stability of models.
Our research aims to develop new sample selection techniques that will increase the robustness of models to noise and adversarial attacks, improve learning, and provide a framework for explaining learning processes. The results of the project will contribute to the development of reliable and trustworthy AI systems, in line with the European Union's strategic goals for “Explainable and Robust AI.”
The preliminary results of our research are promising. We have developed a novel streaming approach with Base-Value Mechanism that significantly improves the training process of neural networks. Our proprietary method, called Persistent Entropy-Based (PEB), uses base values to indicate the importance of individual training elements. In tests on the EMNIST dataset, the method achieved higher accuracy on test data compared to traditional methods, confirming the effectiveness of our approach.
The project has great potential to bring important innovations to the field of AI, contributing to more resilient and understandable machine learning models that will better serve society in a variety of applications.
The project is coordinated by Mateusz Wojtulewicz, PhD student supervised by Prof. Leszek Rutkowski.
The project is funded by the National Science Centre under the Agreement No. UMO-2024/53/N/ST6/02844 z 5.02.2025 r.
