Abstract
Humans can reason at an abstract level and structure information into abstract categories, but the underlying neural processes have remained unknown. Recent experimental data provide the hint that this is likely to involve specific subareas of the brain from which structural information can be decoded. Based on this data, we introduce the concept of assembly projections, a general principle for attaching structural information to content in generic networks of spiking neurons. According to the assembly projections principle, structure-encoding assemblies emerge and are dynamically attached to content representations through Hebbian plasticity mechanisms. This model provides the basis for explaining a number of experimental data and provides a basis for modeling abstract computational operations of the brain.
Introduction
One of the most enduring mysteries in cognitive science is how the biological brain performs "variable binding"—the ability to flexibly link abstract roles (like "Agent" or "Patient") to specific sensory content (like "truck" or "ball"). While modern AI excels at pattern recognition, it often struggles with compositionality: understanding that "the cat chased the dog" is structurally identical to "the dog chased the cat," despite the roles being reversed. This paper, published in eNeuro, presents a biologically plausible framework termed "assembly projections" to bridge the gap between low-level neural dynamics and high-level symbolic computation.
The Assembly Projection Mechanism
The researchers propose a model consisting of a content space (representing concepts) and several neural spaces (representing structural roles). The system utilizes stochastically spiking neural networks (SNNs) with realistic membrane potentials and refractory periods.
The core of the model is the "projection" process. When a specific concept assembly is active in the content space and a structural neural space is disinhibited (the "inhibitory lock" is lifted), a new assembly emerges in that neural space. Through Spike-Timing-Dependent Plasticity (STDP), the network learns the link between a concept and a role in real-time. This allows the brain to use disinhibition as a gating mechanism to control when and where information is "written" or "read" from these neural registers, similar to operations in a CPU.
Key Insights: Factorization and Randomness
A critical finding is that these projections emerge naturally from random, sparse connectivity. A perfectly pre-wired brain is not required for complex logic; rather, plasticity and a gating mechanism are sufficient. This leads to factorized coding, where the brain represents "truck" and "agent" separately and links them dynamically. This factorization provides the infinite flexibility characteristic of human thought. Furthermore, by using distributed assemblies rather than single-neuron pointers, the model remains robust to noise and individual neuron loss, ensuring biological realism.
Experimental Validation and Results
The model's performance was validated against synthetic linguistic patterns and experimental fMRI data.
- High Recall Accuracy: The model achieved a readout error of only 0.9% ± 0.2% when retrieving content from structural roles after a delay.
- Robustness: Performance remained high even when synaptic plasticity parameters were varied by up to 20%, suggesting the mechanism is a fundamental property of the architecture.
- Decoding Sentence Meaning: Classifiers trained on simulated neural activity could distinguish between sentences like "The truck hit the ball" and "The ball hit the truck" with 0% error on held-out data. This mirrors the decoding capabilities found in human fMRI studies of the mid-superior temporal cortex (lmSTC) by Frankland and Greene (2015).
- Symbolic Operations: The authors demonstrated that the model could perform "atomic" computational tasks, such as copying information between roles and comparing content across different registers.
Limitations and Future Directions
While promising, the model currently handles a limited vocabulary (up to 40 concepts) and "flat" structures. Future work is required to determine if "stacking" these projections can represent complex, nested hierarchical sentences, addressing the MERGE operation central to linguistics. Additionally, the mechanism for "executive control"—the signal that decides when to disinhibit a specific neural space—remains an open question, as the current model assumes an external controller.
This summary is based on research published in eNeuro.
Summary updated