At its core, Albatross is a research project in the area of continual learning.
Continual Learning is an approach in artificial intelligence and machine learning where models learn and adapt continuously from new data over time. Unlike traditional models trained on a fixed dataset, continual learning models update their knowledge incrementally, allowing them to adapt to new tasks and environments without forgetting previously learned information. This approach aims to avoid catastrophic forgetting, efficiently utilize new data, and maintain the ability to generalize across different contexts, making the model more robust and flexible in dynamic and ever-changing environments.
As a research project, Albatross aims to achieve the following main goals:
Leverage differences between classical deep learning and lifelong learning approaches to develop innovative training techniques that support efficient static training and robust, long-term lifelong learning. Design adaptable model architectures, focusing on modular networks and timeless knowledge structures to consolidate information despite changes in model or data distribution. Identify scalable data representations, using synthetic data generation and evolving features to minimize forgetting and computational costs. Create a comprehensive benchmark for lifelong learning, analogous to the ``ImageNet for Lifelong Learning,’’ to evaluate the proposed techniques in a realistic long-term input dynamic context and drive the whole field forward.
