Assembly projections support the assignment of thematic roles to concepts in networks of spiking neurons

Robert Legenstein, Michael G. Müller, Christos H. Papadimitriou, Santosh Vempala, Wolfgang Maass

Abstract

Currently existing neural network models lack an important computational capability of networks of neurons in the brain: the capability to reason on an abstract level, and to structure information according to abstract categories. While hardly anything is known about the processes underlying these capabilities in the human brain, experimental evidence for the organization of an important prerequisite for abstract reasoning has recently been found: for assigning words to thematic roles, such as subject or object. Such data from fMRI recordings indicate that assemblies in subregions of temporal cortex which encode thematic roles are shaped by concepts that are assigned to that role. We propose a model for the assignment of thematic content to variables that is consistent with these data and provides new functions to neural network computations. We provide a proof of principle that this model can be implemented even in networks of spiking neurons. It enables structured information retrieval and copying of information without routing, and provides a basis for the emulation of more demanding cognitive computations in networks of spiking neurons.

Introduction: The Neural Binding Problem

A fundamental challenge in neuroscience and artificial intelligence is the "binding problem": how does the brain represent structured relationships between concepts? For instance, to distinguish between "The truck hit the ball" and "The ball hit the truck," the brain must bind specific concepts (truck, ball) to thematic roles (agent, patient). While traditional computer architectures use registers and pointers, biological neural networks must rely on plastic, distributed representations.

Recent fMRI studies by Frankland and Greene (2015) identified regions in the left mid-superior temporal cortex (lmSTC) that appear to encode these roles. Crucially, these representations are content-specific; the neural signature of an "agent" changes depending on whether the agent is a truck or a ball. This paper proposes a biologically plausible mechanism—Assembly Projections—to explain how these structured representations emerge through learning in spiking neural networks.

Architecture and Binding Mechanism

The proposed model consists of two primary components: a Content Space ($\\mathcal{C}$) representing concepts (e.g., "concept cells" for words) and multiple Variable Spaces ($\\mathcal{V}$) representing thematic roles. These spaces are connected by sparse, potential feedforward and feedback connections.

Architecture of the assembly projection model showing the binding of concepts to variable spaces via disinhibition and Hebbian plasticity.

The binding process relies on three key biological mechanisms:

  1. Disinhibition as a Gate: Variable spaces are normally inhibited. To bind a concept to a role, a specific variable space is disinhibited (opened).
  2. Hebbian Plasticity (STDP): When a concept assembly is active in $\\mathcal{C}$ and a variable space $\\mathcal{V}$ is disinhibited, Spike-Timing-Dependent Plasticity (STDP) strengthens the connections between the active neurons. This creates a new, sparse assembly in $\\mathcal{V}$ that serves as a "pointer" to the content in $\\mathcal{C}$.
  3. Transient Excitability: Short-term memory is maintained through a temporary boost in neuronal excitability, allowing the network to reactivate the most recently formed pointers during recall.

Information Retrieval and Spiking Dynamics

The model demonstrates high fidelity in retrieving bound information. During a "RECALL" phase, the variable space is disinhibited again. Due to the strengthened synaptic connections and transient excitability, the specific assembly in $\\mathcal{V}$ reactivates, which in turn triggers the original concept assembly in $\\mathcal{C}$ via feedback connections.

Temporal dynamics of the CREATE and RECALL operations, illustrating spiking patterns in content and variable spaces.

In simulations involving 250 trials, the model achieved a 100% success rate in retrieving the correct concept. Furthermore, the model's internal representations mirror human fMRI data. A linear classifier trained on the Variable Spaces could distinguish sentence meanings (e.g., who hit whom) with only a 3% error rate, whereas a classifier trained on the Content Space alone performed near chance (44% error), consistent with the findings of Frankland and Greene.

Cognitive Primitives and Symbolic Computation

Beyond simple binding, the assembly projection framework supports "atomic" cognitive operations:

  • COPY: Moving a value from one variable to another by activating the source variable and disinhibiting the target variable.
  • COMPARE: Determining if two variables hold the same content by checking for co-activation or interference between their respective assemblies.

These operations provide a bridge between the messy dynamics of spiking neurons and the structured, symbolic-like processing required for higher-order cognition.

Conclusion and Future Directions

This research provides a proof-of-principle for how spiking neurons can implement variable binding without rigid, hard-wired routing. By using distributed assemblies as pointers, the system is robust to noise and biologically realistic.

Future work will address the "controller" problem—how the brain decides which variables to disinhibit—and explore scaling the model to larger vocabularies and more complex syntactic structures, such as nested clauses, potentially through linguistic operations like "MERGE."

Summary updated

Abstract

Currently existing neural network models lack an important computational capability of networks of neurons in the brain: the capability to reason on an abstract level, and to structure information according to abstract categories. While hardly anything is known about the processes underlying these capabilities in the human brain, experimental evidence for the organization of an important prerequisite for abstract reasoning has recently been found: for assigning words to thematic roles, such as subject or object. Such data from fMRI recordings indicate that assemblies in subregions of temporal cortex which encode thematic roles are shaped by concepts that are assigned to that role. We propose a model for the assignment of thematic content to variables that is consistent with these data and provides new functions to neural network computations. We provide a proof of principle that this model can be implemented even in networks of spiking neurons. It enables structured information retrieval and copying of information without routing, and provides a basis for the emulation of more demanding cognitive computations in networks of spiking neurons.
BibTeX
@inproceedings{legenstein2020assembly,
  author = {Robert Legenstein and Michael G. Müller and Christos H. Papadimitriou and Santosh Vempala and Wolfgang Maass},
  title = {{Assembly projections support the assignment of thematic roles to concepts in networks of spiking neurons}},
  booktitle = {eNeuro},
  year = {2020},
  doi = {10.1523/eneuro.0533-19.2020},
}

Loading related content...

Figure