Neural Assemblies Can Learn Causal Direction, Marking a Key Step Forward for Brain-Inspired Computing

Introduction: When Causal Reasoning Meets Brain-Inspired Computing

Causal reasoning has long been one of the most challenging core problems in artificial intelligence. The human brain seems to possess an innate ability to determine the direction between "cause" and "effect," but how exactly is this ability realized at the neural level? A recent interdisciplinary paper published on arXiv (arXiv:2604.26919v1) offers an exciting answer: Neural Assemblies — groups of neurons that fire together and strengthen their connections through co-activation — can learn the direction of causal influence between variables.

This discovery deeply integrates two cutting-edge fields — brain-inspired computing and causal learning — providing an important theoretical foundation for understanding the brain's causal reasoning mechanisms and building more intelligent AI systems.

Core Finding: Inherent Operations of Neural Assemblies Are Sufficient for Causal Learning

Neural Assemblies are a brain-inspired computational model whose core idea derives from the classic Hebbian learning theory — "neurons that fire together wire together." Previous research has demonstrated that Neural Assemblies, as a computationally universal substrate, can accomplish complex cognitive tasks such as classification, syntactic parsing, and planning. However, a critical question has remained unresolved: Can Neural Assemblies internalize causal directionality?

Through rigorous theoretical analysis and experimental validation, the research team proved that the inherent operations of Neural Assemblies — including Projection and Local Plasticity Control — are sufficient for the system to autonomously learn and encode the direction of causal relationships. In other words, without introducing any additional complex mechanisms, Neural Assemblies inherently possess the ability to discover whether "A causes B" or "B causes A."

This conclusion carries profound implications. It suggests that causal reasoning may not require some high-level symbolic reasoning module, but could instead be a capability that naturally emerges at the fundamental level of neural networks.

Technical Breakdown: Why Can Neural Assemblies Capture Causal Direction?

To understand this achievement, it is necessary to review several core operations of Neural Assemblies:

Projection: When a group of neurons (Assembly A) sends signals to another brain region, it activates and gradually forms a new stable assembly (Assembly B) in the target area. This process inherently carries directionality — signals flow from source to target, closely aligning with the temporal structure of "cause precedes effect" in causal relationships.

Local Plasticity Control: Synaptic strengths within Neural Assemblies are dynamically adjusted based on local activity. This mechanism enables the system to distinguish causal direction based on statistical asymmetries in observational data. In causal relationships, the statistical dependencies between cause and effect are typically asymmetric, and local plasticity is precisely suited to capture this asymmetry.

Co-activation and Competition: The co-activation within Neural Assemblies and the competitive mechanisms between assemblies provide the system with a natural way to represent variables, allowing causal relationships to be explicitly encoded in the connection patterns between assemblies.

From a computational perspective, the combination of these operations endows Neural Assemblies with functionality similar to classical causal discovery algorithms (such as the PC algorithm or intervention-based methods), but the implementation is entirely biologically plausible, relying on neither global optimization nor backpropagation.

Comparison with Existing Methods

Current causal learning in AI primarily relies on two categories of methods: constraint-based methods using statistical independence tests, such as the PC and FCI algorithms; and score-based search methods, such as the GES algorithm. While these methods have rigorous theoretical guarantees, they typically require global reasoning, have high computational complexity, and struggle to explain how the brain accomplishes similar causal reasoning tasks.

By contrast, the Neural Assembly model offers several unique advantages:

• Biological Plausibility: All operations are based on known neuroscience mechanisms, requiring neither backpropagation nor global loss functions
• Local Computation: Causal direction judgments are made entirely through local synaptic updates, possessing inherent distributed and parallel characteristics
• Online Learning: The system can progressively update its causal model as data streams in, without requiring batch processing
• Scalability: The Neural Assembly model naturally supports large-scale parallel computation and is well-suited for deployment on neuromorphic hardware

Impact and Outlook

The significance of this research extends far beyond the academic realm. It opens new possibilities in multiple directions:

Cognitive Science: The research provides a computational-level hypothesis for the long-standing question of "how the brain performs causal reasoning." If causal learning can indeed be achieved through the basic operations of Neural Assemblies, then causal cognition may be a fundamental capability of the brain rather than an advanced function acquired through learning.

Neuromorphic Computing: With the development of neuromorphic chips such as Intel Loihi and IBM TrueNorth, causal learning algorithms based on Neural Assemblies could potentially be deployed directly on low-power hardware, enabling efficient edge-based causal reasoning.

AI System Design: Current large language models still exhibit notable shortcomings in causal reasoning, frequently confusing correlation with causation. Incorporating the causal learning mechanisms of Neural Assemblies into AI system design could fundamentally enhance models' causal reasoning capabilities.

Human-Machine Interaction and Decision Support: AI systems equipped with causal reasoning capabilities can better understand the effects of interventions, providing more reliable recommendations in scenarios requiring causal judgment, such as medical diagnosis and policy-making.

Of course, this work is still in the theoretical exploration stage, and there is a long road from the models described in the paper to practical applications. Future work will need to validate the method's effectiveness in more complex causal structures, high-dimensional data, and real-world scenarios involving confounding factors. Additionally, how to effectively integrate Neural Assembly-based causal learning with existing deep learning frameworks is a direction worthy of in-depth investigation.

Nevertheless, this paper presents an exciting possibility: causal reasoning may not require complex symbolic systems or elaborate algorithmic designs — the most fundamental neural computational mechanisms in the brain already contain the power to discover causal relationships. This is not only an important expansion of the boundaries of brain-inspired computing capabilities, but also a critical step toward truly understanding the nature of intelligence.