PaperlayerView exact revision on arXiv

Source-linked AI summary

Rate-Coding Bundle Memory: A Unified Model of Memory and Control for Symbolic Computation in the Brain

Teun van Gils, Rowan P. Sommers, Markus Ostarek, Peter Hagoort

arXiv:2608.29189v1q-bio.NCcs.AIcs.CLcs.SC

TL;DR

The paper addresses how cognition can combine continuous, graded neural representations with discrete symbolic operations. It develops RCBM, a neurobiologically plausible hybrid model integrating rate-coded semantics, bundle memory, and control, and reports near-perfect performance across diverse S3R tasks. The model is offered as a proof-of-concept for the Symbolic Subsystem Hypothesis, while its current dependence on bundle memory and lack of hierarchical compositionality remain limitations.

  • Problem

    Purely connectionist systems face challenges with one-shot learning, pattern separation, compositionality, and controlled memory operations, motivating a hybrid account of cognition.

  • Method

    RCBM embeds a symbolic subsystem within a rate-coding connectionist architecture combining semantic processing, bundle memory, and neurobiologically plausible control mechanisms.

  • Results

    RCBM achieves near-perfect performance across a diverse set of S3R tasks, while lesion studies identify control and memory subsystems as causally necessary for explaining these phenomena.

  • Takeaways & Limitations

    RCBM provides a proof-of-concept that neural mechanisms can support both continuous semantic representation and discrete symbolic operations across a broad range of cognitive phenomena.

  • Takeaways & Limitations

    The model lacks hierarchical compositionality and may be over-reliant on bundle memory, motivating a more robust cortical short-term memory system.

Abstract

from arXiv · show

We propose a neurobiologically plausible model of cognition that combines the advantages of connectionist and symbolic systems, and that can explain a wide range of cognitive phenomena. This model, called Rate-Coding Bundle Memory (RCBM), is based on the Symbolic Subsystem Hypothesis, which posits that the brain implements a symbolic subsystem within its fundamentally connectionist nature. RCBM is a hybrid model that uses rate coding to represent symbols in a continuous space, and it uses a bundle memory system to store and retrieve these symbols. The model is capable of solving a wide range of cognitive phenomena, including one-shot learning, pattern separation, and the binding problem. We argue that RCBM provides a promising framework for understanding the nature of cognition, and that it can be used to develop more sophisticated models of cognition in the future.

Introduction

The paper frames cognition as requiring both graded, context-sensitive representations and discrete symbolic operations. It proposes the Symbolic SubSystem Hypothesis and the S3R phenomena as a framework for evaluating a neurobiologically plausible hybrid model.

  • Motivation: Human categories have fuzzy, context-dependent boundaries that resist rigid definitions.The paper uses birds as an example of concepts whose membership depends on context and goals.
  • Motivation: Symbolic cognition is motivated by the ability to combine discrete symbols into complex, rule-governed ideas.The paper distinguishes flat compositionality from hierarchical recursive compositionality.
  • Motivation: One-shot learning, pattern separation, and compositionality are difficult for purely connectionist systems but tractable through symbolic operations such as variable assignment.Connectionist systems often require repeated examples, whereas symbolic systems can assign new variables after a single observation.
  • Proposal: RCBM is presented as a principled hybrid approach intended to explain a broad set of Symbol Recombination, Retention, and Resolution phenomena.The model aims to combine the advantages of connectionist and symbolic traditions within one neurobiologically plausible system.
  • Motivation: Cognitive memory also requires controlled retrieval, suppression, disambiguation, and switching between operations to prevent interference.The paper identifies control as necessary for managing overlapping memories and task-dependent operations.
  • Proposal: The Symbolic SubSystem Hypothesis proposes that a connectionist brain implements a symbolic subsystem that works alongside broader neural processing.The hypothesis is deliberately general about the subsystem’s location and task coverage.

Methods

RCBM embeds semantic representation, bundle memory, and control in one rate-coding connectionist system. Its mechanisms encode, retrieve, disambiguate, protect, and clear memory bundles according to input and task context.

  • Methods: RCBM implements the prior Bundle Memory model as a single rate-coding connectionist system with neurobiologically plausible memory and control mechanisms.The model retains the semantic, bundle-memory, and control components while embedding them in one system.
  • Methods: The semantic system uses hierarchically organized attribute neurons connected to visual and auditory sensory neurons and linguistic inputs.Attributes can be activated externally through sensory or linguistic input.
  • Methods: The bundle-memory system contains Winner-Take-All Memory and Multiple Activation Memory pools for storing and managing discrete memory bundles.Bundles are represented by paired WTAM-MAM neurons within a discretized neural state space.
  • Methods: Control neurons detect novelty, existing matches, multiple candidates, or suppression requirements and regulate memory allocation and retrieval.They allocate new bundles, protect existing memories, activate appropriate bundles, and clear memory between sequences when needed.
  • Methods: Novel inputs receive new WTAM slots, matching inputs retrieve existing bundles, and ambiguous inputs trigger the MAM pool until disambiguation.Semantic information is reactivated from the active bundle and newly encountered information is bound to memory neurons.

Rate Coding Model

RCBM uses rate coding to model neural activity while reducing the computational demands of spiking simulations. Its semantic, control, and memory components process inputs and retain or retrieve bundles over time.

  • Rate Coding Model: Rate coding represents information through neuronal firing rates while reducing computational requirements relative to spiking neural networks.The model clips firing rates between 0 and 1 and uses matrix-based signal propagation.
  • Rate Coding Model: Leaky neurons update their state from prior activity, incoming signals, decay toward baseline, and optional noise.The update uses the activation function f, baseline b, incoming weight matrix W, decay rate d_r, and noise ζ.
  • Rate Coding Model: The semantic system contains sensory and conceptual feature neurons that can receive visual, auditory, or linguistic input.Conceptual features are organized hierarchically, and words in the artificial language activate corresponding conceptual feature neurons.
  • Rate Coding Model: The control and memory systems coordinate novel-memory allocation, existing-memory retrieval, suppression, and bundle manipulation.The model uses WTAM and MAM memory pools connected to conceptual feature neurons and control neurons.
  • Rate Coding Model: During an example sequence, novel input creates and binds a bundle, an end-of-sentence signal suppresses memory, and repeated input reactivates it.The sequence proceeds through distinct time windows from 0 ms to 3500 ms.

Conjunction Neurons as the Basis of Bundle Memory

Bundle Memory builds on conjunction-neuron mechanisms and Fast Hebbian Plasticity to create dynamic, bidirectional bindings between memory neurons and semantic attributes.

  • Conjunction Neurons as the Basis of Bundle Memory: Fast Hebbian Plasticity increases connection strength when source and target neurons are active together, enabling dynamic memory traces.The model uses specialized FHP functions and excludes self-connections.
  • Conjunction Neurons as the Basis of Bundle Memory: Conjunction neurons provide the basis for dynamically binding memory representations to attribute neurons.The mechanism resembles prior conjunction-neuron models of working memory.
  • Conjunction Neurons as the Basis of Bundle Memory: Bidirectional facilitating connections let memory neurons reactivate attributes and let attributes retrieve associated memories.Connections from memory neurons to attributes are stronger than the reverse connections.
  • Conjunction Neurons as the Basis of Bundle Memory: Lateral inhibition between memory neurons creates a winner-take-all mechanism for selecting one bundle.A negative scaling factor between memory neurons suppresses competing memory neurons.
  • Conjunction Neurons as the Basis of Bundle Memory: Noise selects an unassigned WTAM neuron when no distinguishing attributes identify a novel memory trace.This permits random assignment of unassigned WTAM neurons to new bundle memories.

Increasing Network Stability

RCBM increases memory stability and discreteness through bistable neurons, decaying Hebbian weights, thresholded binding, and adaptive lateral inhibition.

  • Increasing Network Stability: The extended model addresses complex S3R phenomena that the earlier conjunction-neuron model struggled to explain.The motivating examples include discourse incrementation and the problem of two.
  • Increasing Network Stability: Weight decay gradually forgets memory traces, preventing saturation and leaving capacity for new traces over time.Weights decrease toward a specified baseline at rate d_w, rather than relying only on active unlearning.
  • Increasing Network Stability: Non-zero attribute-to-memory baselines allow unbound attributes to activate memory neurons and form new traces.Memory-to-attribute connections can use a zero baseline.
  • Increasing Network Stability: Sigmoidal synaptic thresholds restrict Hebbian binding to cases with strong pre- and post-synaptic activation.This supports the binary interpretation that attributes are either included in a bundle or not.
  • Increasing Network Stability: Bistable neurons provide upper and lower attractor states, making memory neurons more stable and discrete.They strengthen the winner-take-all dynamic so that only one memory neuron is activated at a time.
  • Increasing Network Stability: Adaptive lateral inhibition balances memory allocation against retrieval by starting weak and increasing with consistent coactivation.Weak initial inhibition reduces noise interference, while later strengthening accelerates competition for new memory slots.

Control System

The control system manages memory allocation and retrieval, while a parallel MAM pool detects ambiguity without winner-take-all suppression.

  • Control System: The control system distinguishes novel, existing, ambiguous, and suppressed states to manage memory operations.Its control neurons allocate new bundles, protect existing ones, detect competing candidates, and suppress the memory system when needed.
  • Control System: Novelty detection boosts unassigned WTAM neurons when activated attributes are not associated with an existing bundle.Existing-memory detection instead boosts assigned memory neurons and prevents unassigned neurons from competing.
  • Control System: Memory suppression turns off WTAM neurons so semantic activity can occur without immediate bundle binding or activation.This separates semantic processing from memory operations when required.
  • Control System: The MAM pool maps one-to-one onto WTAM neurons but lacks lateral inhibition, allowing multiple candidate memories to remain active.MAM neurons receive direct attribute input but do not reactivate attributes, preventing cross-bundle attribute mixing.
  • Control System: Ambiguity detection inhibits the WTAM pool and postpones winner-take-all resolution until disambiguating information arrives.The ambiguity neuron responds when multiple MAM neurons are simultaneously active.
  • Control System: MAM synchrony and Hebbian plasticity stabilize the simultaneous activation needed to detect competing bundles.Activation differences are thresholded into synchrony values that modulate FHP.

Semantic Network

The model uses a deliberately simplified artificial language and a separate semantic network to connect multimodal attributes with bundle memory. This design isolates memory and binding phenomena from natural-language complexity.

  • Artificial language: The language omits features of natural languages, including subject-object distinctions and part-of-speech information.This simplification prevents the model from implementing additional syntactic and semantic complexities.
  • Artificial language: The artificial language maps each word one-to-one onto a semantic attribute rather than encoding grammatical categories.The model treats apparent nouns, adjectives, and verbs as attributes distinguished by semantic content.
  • Semantic architecture: The semantic system contains visual and auditory input layers with attribute classes that feed a concrete-semantic hidden layer in a Hub-and-Spoke-like architecture.The hidden layer forms the first layer of the semantic hub.
  • Semantic architecture: The semantic network has 2 modalities, 28 attributes, and 96 neurons, while remaining separate from the bundle-memory system.Memory interfaces with the semantic network, but the semantic information is not copied into the memory system.

Model Parameters

The model’s parameters were manually tuned for stability and performance across a broad S3R test suite, with mechanisms intended to reduce sensitivity to exact parameter choices. The evaluation covers storage, retrieval, interleaved memories, and semantic association.

  • Parameter selection: Parameters were manually tuned from prior settings to balance network stability with performance across many S3R phenomena.The authors state that parameter interactions were analyzed in some cases, including lateral inhibition strength and noise level.
  • Evaluation suite: The model evaluates 14 tasks across controlled storage, memory retrieval, interleaved memories, and semantic association.Tests compare model outputs with expected outputs and assess whether the full model performs with high accuracy.
  • Semantic association: The tasks probe semantic association through deduced association, correlation inference, and sensory binding.These tests examine associations among attributes, memory references, and links between sensory and conceptual representations.
  • Controlled storage: Controlled-storage tasks test correlation violations and both inferred and explicitly determined novelty.They assess storing atypical attribute combinations and allocating new memory neurons for novel information.
  • Memory retrieval: Memory-retrieval tasks assess gradual forgetting, long-term remembering, memory maintenance, and pattern completion.Memory maintenance specifically targets binding and one-shot learning, while pattern completion uses partial input to retrieve a learned sequence.
  • Interleaved memories: Interleaved-memory tasks assess competition, incremental binding, multiple bundles, and pattern completion after a distractor sequence.The many-memories task involves retrieving up to four bundles without interference.

Lesioning the Model

The lesion analysis removes selected network connections while preserving the neurons and remaining connections, enabling side-by-side tests of network mechanisms. Each test is rerun across lesion conditions for around 12000 trials.

  • Lesion design: Lesions selectively remove connections associated with memory detection, ambiguity, novelty, suppression, sensory grounding, or winner-take-all completion.The lesion design leaves neurons and other connections intact for comparative evaluation.
  • Memory mechanisms: The MAM lesion removes WTAM-to-MAM and MAM-to-attribute connections, while the MAM-synchrony lesion removes lateral facilitatory MAM connections.These lesions target distinct parts of the memory allocation and synchrony mechanisms.
  • Semantic mechanisms: The sensory-grounding lesion removes connections between semantic layers, leading to loss of semantic information.This condition tests the role of semantic grounding in model function.
  • Memory mechanisms: The WTAM lesion removes bidirectional connections between attribute neurons and WTAM neurons.This lesion isolates the contribution of the attribute-WTAM interface.
  • Evaluation: Around 12000 trials were conducted across the tests and lesioned network versions, in addition to the full unlesioned model.The same tests were rerun for each specified lesion condition.

Tools

The model was implemented in Python with Aesara for simulation optimization and execution, alongside standard numerical, data-handling, visualization, and statistical-analysis tools.

  • Implementation: RCBM was implemented in Python 3.10.9 using Aesara 2.9.3 to optimize and run simulations.NumPy and Pandas supported data handling and aggregation, while Matplotlib supported visualization.
  • Analysis: Simulation data were analyzed in R 4.2.1 using Tidyverse 2.0.0 for data processing.The implementation also used Matplotlib 3.6.2 for data visualization.

Results

RCBM was evaluated on controlled tasks targeting Symbol Recombination, Retention, and Resolution, including memory storage, retrieval, ambiguity resolution, and incremental binding. It achieved over 95% performance across tasks, while lesions produced systematic impairments.

  • Evaluation design: The evaluation tested RCBM’s ability to store, retrieve, and manipulate information relevant to linguistic memory construction.The task set targeted key challenges in cognitive modeling related to Symbol Recombination, Retention, and Resolution.
  • Overall performance: >95% performance was achieved on all tested S3R tasks.The model performed successfully across the broad task set shown in Figure 4.
  • Controlled storage: RCBM identified novel entities, created new memory neurons when required, and stored anti-correlated attributes for correct retrieval.These capabilities were demonstrated through inferred novelty, determined novelty, and correlation-violation tests.
  • Memory retrieval: The model retained coherent memories, completed partial patterns, gradually forgot old information, and handled interleaved memories with overlapping attributes.It also incrementally bound information to entities and created up to four memory bundles in varied sequence patterns.
  • Mechanisms: The control system uses Hebbian learning, lateral inhibition, and facilitating synapses to allocate memory, detect novelty and ambiguity, and suppress irrelevant activity.These mechanisms are embedded within a single rate-coding framework.
  • Lesion analysis: Lesions caused systematic performance reductions, revealing impairments in disambiguation, semantic association, memory capacity, or complete functioning.The lesion study was used to investigate which connections and mechanisms were required for model functioning.

Discussion

The discussion presents RCBM as a neurobiologically plausible connectionist implementation of symbolic cognition, supported by controlled memory mechanisms and dynamic binding. It also identifies dependence on bundle memory and unconfirmed synaptic predictions as important boundaries.

  • Contribution: RCBM demonstrates that symbolic phenomena can be modeled in a connectionist and neurobiologically plausible manner.The model embeds control mechanisms based on known neural processes within a single rate-coding framework.
  • Control system: The control system detects novelty, ambiguity, and memory states while suppressing memory activity when necessary.Its mechanisms include lateral inhibition and specialized facilitating synapses.
  • Findings: The lesion study found near-perfect task performance overall and indicated causal importance for specific control and memory components.The findings position RCBM as a proof-of-concept for the Symbolic Subsystem Hypothesis.
  • Limitations: The model may be over-reliant on bundle memory because lesioning the novelty control neuron prevents new information from being stored there.The authors suggest adding a more robust cortical short-term memory system to reduce this dependence.
  • Neurobiological scope: Neurobiological plausibility does not establish that the model’s mechanisms are actually observed in the brain.The proposed symbolic subsystem may be distributed across multiple regions rather than localized to one structure.
  • Binding mechanisms: The model predicts reciprocal attribute–bundle connections and selective strengthening through compartmentalized synapses.More evidence is required to confirm that such synapses bind semantic attributes only to particular memory neurons.
  • Binding mechanisms: Bundle memory supports dynamic variable binding, although the current RCBM implementation does not multiplex multiple distinct locations in one context.The discussion connects this capability to context-dependent remapping and feature binding in hippocampal place cells.

Distinction between the WTAM and MAM pools

RCBM separates memory retrieval into WTAM and MAM pools to balance interference avoidance with the ability to retain multiple similar candidates. This division supports controlled retrieval while predicting parallels with hippocampal DG and CA3 circuitry.

  • WTAM pool: WTAM activates only one bundle at a time, preventing interference between memories with overlapping or contrasting attributes.This parallels sparse coding and lateral inhibition associated with dentate gyrus function.
  • Interaction between pools: When competing candidates are similarly relevant, suspending winner-take-all selection allows both to remain active until additional information resolves the ambiguity.A WTA-only mechanism would otherwise choose arbitrarily between equally plausible candidates.
  • MAM pool: MAM keeps multiple similar memories active simultaneously through auto-associative binding without mixing their semantic content.The model links this function to recurrent excitatory connectivity in CA3.
  • Neurobiological prediction: RCBM predicts a division of labor between hippocampal DG and CA3 analogous to WTAM and MAM, while allowing analogous circuits elsewhere, such as prefrontal cortex.The prediction concerns DG-mediated separation and CA3-like auto-associative maintenance.
  • Controlled sequential retrieval: The model retrieves one bundle at a time and uses suppression to reinitialize competition so a different bundle can become active.Currently, new sensory input triggers retrieval after suppression, while novelty and ambiguity provide additional control signals.
  • Controlled sequential retrieval: RCBM does not yet directly activate another bundle after suppression, limiting active searching or iteration over bundles.The authors identify this as a missing control mechanism rather than a failure of the WTAM–MAM distinction.

Localization of control signals

RCBM locates control in mechanisms that interact recurrently with semantic processing rather than in an isolated central executive. The model relates novelty and ambiguity signals to memory regulation and proposes distributed neurobiological implementation through neuromodulatory and circuit mechanisms.

  • Control signals: Novelty and ambiguity signals are predicted to control memory encoding and retrieval, not merely to be detectable correlates of processing.The model connects novelty to newly arrived information and ambiguity to interference between similar memories.
  • Control signals: Brain evidence indicates that novelty and ambiguity can be detected and used to affect memory encoding, supporting a memory-control system with signals resembling those in RCBM.The authors cite novelty-related ERPs, ambiguity-sensitive Nref responses, and cognitive-control evidence against similarity-based interference.
  • Neuromodulatory implementation: Dopamine and norepinephrine are proposed candidates for novelty signaling, while acetylcholine is proposed for novelty and ambiguity control.Acetylcholine may increase dentate-gyrus lateral inhibition, a mechanism associated with pattern separation.
  • Neuromodulatory implementation: Diffuse neuromodulation resembles RCBM’s novelty node, although the model restricts novelty effects to unbound memory nodes.This is a proposed correspondence, not an identity between neuromodulators and model nodes.
  • Sequential scheduling: Theta-linked cholinergic activity is proposed to separate encoding and retrieval phases, implementing serial scheduling of memory operations similar to RCBM.The hypothesis places encoding at theta peaks and retrieval at troughs.
  • Scope boundary: The authors caution that RCBM’s control system is greatly simplified and is not a literally neurobiologically realistic implementation of a symbolic subsystem.They present it as a source of plausible mechanisms and cortico-hippocampal parallels rather than a literal brain model.
  • Distributed control: The model implements concrete mechanisms for some executive-control operations, addressing concerns that control theories leave neural implementation underspecified.Its primary control claims are tied to mechanisms currently implemented in RCBM.
  • Distributed control: RCBM uses a dedicated top-down control structure that recurrently interacts with bottom-up semantic processing, allowing control to depend on novelty or ambiguity in the input.This division of labor separates abstract control from semantic representation without making control independent of semantic processing.

RCBM exemplifies how dynamical systems can be symbolic

RCBM shows how a neural dynamical system can support abstract symbolic operations while remaining a continuous system. Its bundle memories provide controlled retrieval and novelty-based memory formation, but hierarchical compositionality remains unresolved.

  • Dynamical systems and symbolic operations: RCBM is implemented as coupled differential equations while supporting discrete read and write operations over bundle memory.The model can therefore be described both as a symbolic computer and as a continuous dynamical system.
  • Dynamical systems and symbolic operations: Each attribute combination can bind to any bundle, yielding (M + 1) × 2^(M×N) possible attractor states.For M = 6 and N = 28, this corresponds to approximately 10^51 attractor states.
  • Memory control: Attractor dynamics retrieve bound features from existing bundles and detect novelty when no matching memory attractor exists.Novelty detection can trigger formation of a new attractor state.
  • Limitations: RCBM supports flat compositionality but lacks hierarchical compositionality, making relational binding difficult.Flattening relations into multiple bundles would itself require binding between bundles.
  • Hybrid architectures: Bundle memory offers a neurobiologically plausible implementation of binding, unlike some hybrid architectures whose binding operations are not biologically plausible.The comparison concerns the plausibility of the implementation, not symbolic capability alone.

Conclusion

The conclusion presents RCBM as a neurobiologically plausible hybrid architecture that explains several cognitive phenomena through interactions between memory and control. The reported capabilities include one-shot learning, ambiguity resolution, incremental integration, and management of multiple memory traces.

  • Conclusion: RCBM is presented as a neurobiologically plausible instantiation of the Symbolic Subsystem Hypothesis.The model embeds a symbolic subsystem within a fundamentally connectionist framework.
  • Conclusion: RCBM can perform one-shot learning, resolve ambiguity, incrementally integrate information, and manage multiple memory traces.These capabilities are described as key aspects of the S3R phenomena.
  • Conclusion: Lesion studies indicate that these capabilities depend critically on interactions between memory and control subsystems.The conclusion highlights specialized mechanisms for novelty, ambiguity, and memory management.
  • Conclusion: The model’s operations emerge from Hebbian learning, lateral inhibition, and facilitating synapses without abstract or biologically implausible computations.These mechanisms are presented as the neural basis of the reported symbolic operations.
  • Supplementary Materials: The evaluation uses artificial-language stimuli alongside reduced English translations and logically consistent discourses with semantically uncorrelated sentences.The stimuli are designed so recall does not rely on long-term semantic knowledge.
Loading 2608.29189v1…