EARL: Embracing Amnesic Replay for Learning with Noisy Labels
At Axyon AI, we know AI models in live investment environments face a challenge benchmarks rarely capture: data is messy, markets shift, and some signals fed to a model are untrustworthy.
That's why we're proud to share that our recent research, developed with the University of Modena and Reggio Emilia (UNIMORE) and New York University (NYU), has been published in Pattern Recognition, a prestigious peer-reviewed journal in machine intelligence.
The paper introduces EARL, an evolution of our Alternate Experience Replay (AER) framework first presented in "May the Forgetting Be with You" (BMVC 2024). It tackles a key problem in AI: how to keep learning from new data without forgetting what you know or being misled by noisy or incorrect information.
“Modern Deep Neural Networks struggle to retain knowledge in streaming data environments, often leading to forgetting during incremental training. Most Continual Learning approaches address this by rehearsing past data — stored in a replay buffer — while acquiring new knowledge. However, in practical scenarios, noisy labels can contaminate the replay buffer, undermining performance.“
EARL introduces a detailed analysis of learning dynamics in the presence of noise, a more robust evaluation under realistic noise conditions, experiments with pre-trained backbones and modern prompt-based architectures, a study of different sampling strategies, and new benchmarks in Natural Language Processing.
This study addresses this by identifying and filtering noisy data through principled analysis of learning dynamics, turning what is usually seen as a weakness of neural networks, forgetting, into an intelligent tool. The result is a more robust, adaptive model that maintains strong performance under realistic, imperfect conditions.
That conclusion sits at the heart of what we do at Axyon AI: building AI that is not just accurate in controlled settings, but genuinely resilient in real-world applications.
Congratulations to the research team involved: Monica Millunzi, Lorenzo Bonicelli, Angelo Porrello, Jacopo Credi, Petter N. Kolm, and Simone Calderara.