NeuroAI and Beyond: Bridging Between Advances in Neuroscience and ArtificialIntelligence

A landmark paper published on arXiv by 31 leading neuroscientists and AI researchers, including Anthony Zador, Terrence Sejnowski, and Yiannis Aloimonos, lays out a comprehensive roadmap for NeuroAI—artificial intelligence informed by neuroscience. Based on an August 2025 workshop convened by the National Science Foundation, the paper identifies three fundamental capability gaps in current AI: an inability to interact with the physical world (lack of embodiment), inadequate learning that produces brittle systems (poor generalization and robustness), and unsustainable energy and data inefficiency (massive compute and data requirements). The authors argue that bridging these gaps requires a paradigm shift grounded in neuroscience principles.

The roadmap proposes five key neuroscience principles to address these gaps: co-design of body and controller (embodied cognition), prediction through interaction (active inference), multi-scale learning with neuromodulatory control (brain-inspired plasticity), hierarchical distributed architectures (cortical processing), and sparse event-driven computation (neuromorphic computing). The paper outlines a research agenda across near, mid, and long-term horizons, emphasizing the need for interdisciplinary training, hardware access (e.g., neuromorphic chips), community standards, and ethical guidelines. The authors conclude that NeuroAI can overcome current AI limitations while deepening our understanding of biological neural computation, potentially leading to more robust, efficient, and capable AI systems.