Convergent multi-modular architecturefor adaptive learning in Drosophila and artificial intelligence

Faced with dynamic, uncertain environments, a common goal of biological intelligence (BI) and artificial intelligence (AI) is to develop robust adaptive learning capabilities, despite different origins. Exploring shared mechanisms may uncover universal computational principles of information processing. In this perspective, we review recent studies of the <i>Drosophila</i> olfactory learning system, which exemplifies a hierarchical multi-modular architecture with specific connectivity patterns. Through an interdisciplinary comparison of anatomical and functional characteristics, we find that it reflects an elegant combination of two classical multi-modular methods in machine learning: ensemble learning (EL) and mixture-of-expert (MoE). This biological hierarchy incorporates the respective strengths of EL and MoE to improve adaptability and employ effective strategies to address their technical challenges, promoting generalization and alleviating interference on a continual basis. We further propose interdisciplinary research directions, such as developing bio-inspired machine learning models that reconcile EL and MoE, and conducting targeted biological experiments to dissect modular learning functions.