Problem

Current AI systems, despite their impressive performance in specific tasks like image classification, game playing, and natural language processing, achieve these feats by consuming vast amounts of data, computation, and energy, far exceeding what human intelligence requires. They also struggle with generalizing to out-of-distribution data, lack compositional skills, and suffer from catastrophic forgetting. The No Free Lunch Theorem suggests that no single inductive bias can solve all problems, implying a need for problem-specific components. This highlights a critical need for new guiding principles in AI design, leading to the argument that modularity—where specialized components synergistically solve complex tasks—offers a promising solution to these limitations.

Method

This work conducts a comprehensive review across artificial intelligence, neuroscience, evolutionary biology, complex systems, and engineering to establish modularity as a fundamental computational principle. The author defines a modular system as one composed of subsystems with strong internal connections and weak external connections, enabling specialization, robustness, and partial autonomy. The paper formalizes modular AI models using routing and aggregation functions and categorizes modularity in AI into implicit (e.g., layers in DNNs), emergent (e.g., knowledge neurons in LLMs), and architectural (e.g., Mixture-of-Experts, transfer learning adapters). It further explores brain modularity at various scales, from individual neurons and canonical microcircuits to large-scale cortical networks and cognitive abilities, drawing insights from diverse analytical approaches including anatomical studies, fMRI, and psychometric theories. Finally, it contrasts modularity-centered approaches with other brain-inspired AI research.

Results

The review reveals that modularity consistently emerges as a key computational principle across diverse scientific domains. In engineering, modular designs enable scalable and robust systems through operations like splitting, substituting, and porting. In natural systems, hierarchical modularity promotes evolvability, adaptation to changing environments, and efficient information processing. For AI, the paper highlights that deep neural networks inherently possess implicit modularity through their layered structure, which efficiently approximates compositional functions. Emergent modularity, observed in structures like "winning tickets" and "induction heads" in LLMs, demonstrates how specialized components can arise during training. Architecturally, modular designs have proven advantageous in improving generalization and sample efficiency, tackling catastrophic interference in continual learning, boosting positive transfer in transfer learning through adapter modules, and enhancing large language models via Mixture-of-Experts, augmented language models, and multi-modal or neuro-symbolic architectures. Brains, too, exhibit hierarchical modularity at all scales, from the complex processing within individual neurons to the functional specialization of cortical microcircuits and large-scale cognitive networks, all contributing to efficient learning and robust generalization.

Implications

The findings imply that modularity provides a powerful framework for integrating brain organization principles into AI system design to overcome current limitations. By identifying and implementing functional modules analogous to those in the brain—such as dedicated memory and reasoning modules for large language models—AI systems could achieve reduced hallucination rates, improved sample efficiency, enhanced compositional generalization, and lower energy consumption. The paper suggests focusing on the computational level to extract these functional modules from the brain. While acknowledging the success of scaling laws in AI, the author argues that a pure scale-centric approach faces diminishing returns due to increasing costs, environmental impact, and data limitations, emphasizing the continued relevance of brain-inspired inductive biases. Therefore, modularity offers a viable path to developing more efficient, robust, and generally intelligent artificial systems by leveraging nature's evolutionary wisdom.