Causal Learning with Neural Assemblies | AI Research

This paper explores how neural assemblies—groups of neurons that fire together and strengthen through repeated co-activation—can learn the direction of causal influence between variables. While neural assemblies are already known for their ability to classify data and plan actions, this research demonstrates that they can also internalize directed causal relationships. By using a mechanism called DIRECT (DIRectional Edge Coupling/Training), the authors show that neural assemblies can function as an "explainable by design" framework, where causal claims are directly traceable to specific neural activity and synaptic connections.

How the DIRECT Mechanism Works

The core of the framework is the DIRECT mechanism, which teaches the system directed relationships without relying on the complex, opaque optimization methods often used in traditional neural networks. Instead, it uses local plasticity—the brain-inspired process where connections between neurons strengthen based on their activity.
To learn a causal link from a source variable to a target variable, the system co-activates the corresponding neural assemblies while applying a temporary, adaptive increase in "gain" to the forward connection. This process is carefully managed through a "warm-ramp" schedule: it starts with conservative updates to ensure the neural representations are stable, then gradually increases the strength of the directional binding. This ensures that the learned asymmetry between forward and reverse links is a result of the causal training rather than random noise or instability.

Validating Causal Claims

Because the framework is built on local neural dynamics, the researchers can verify their results through two specific, auditable readouts. First, they measure "synaptic-strength asymmetry," which calculates the difference in weight between forward and reverse links. A stronger forward link compared to the reverse link serves as evidence of a learned causal direction.
Second, they use "functional propagation overlap." This test simulates how a signal would flow through the network if only the source assembly were active. By checking if this signal reliably triggers the correct target assembly, the researchers can quantify how well the system has internalized the causal direction. These methods allow the researchers to inspect the "connectome" of the model directly to see exactly how it represents a causal relationship.

Performance and Interpretability

The researchers tested this framework across multiple domains and found that it achieved perfect structural recovery (Precision@K = 1.0) in supervised settings where the ground-truth causal structure was known. By establishing that neural assemblies can preserve and learn causal information, the authors position this approach as a bridge between biologically plausible neural dynamics and formal causal models.
The primary advantage of this approach is its transparency. Unlike backpropagation-based methods, where it is often difficult to trace why a model made a specific decision, this framework allows for "stage-level diagnostic localization." If a causal link is not correctly recovered, the researchers can pinpoint whether the failure occurred during the initial assembly formation, the directional binding phase, or the final readout. This makes the system inherently explainable, as every causal claim is tied to observable physical changes in the neural network.