Attractor and integrator networks in the brain

Mikail Khona; Ila R. Fiete

Source-linked AI summary

Attractor and integrator networks in the brain

Mikail Khona, Ila R. Fiete

arXiv:2112.03978v3Dec 7, 2021q-bio.NC

TL;DR

The review asks how simple neural units can collectively generate persistent, robust representations and computations through attractor dynamics. It synthesizes theoretical mechanisms and experimental validations, concluding that brain circuits construct and use attractor networks while modular reuse can expand flexibility and capacity.

Problem
The review addresses how simple neural elements can support persistent activity, error correction, noisy-cue integration, and flexible representations despite tradeoffs among robustness, capacity, structure, and flexibility.
Method
The review synthesizes attractor-network theory, proposed construction mechanisms, model predictions, and experimental analyses of brain circuits.
Results
The review concludes that the brain constructs and uses attractor networks for several essential computations, including persistent activity, integration, memory, and robust representation.
Takeaways & Limitations
Reusing and recombining modular attractors can produce representations that are robust and structurally constrained while also supporting high capacity and flexibility.
Takeaways & Limitations
Robustness depends on low-dimensional attractor states; a maximally high-dimensional neutrally stable system would lose error correction and noise tolerance.

Abstract

from arXiv · show

In this review, we describe the singular success of attractor neural network models in describing how the brain maintains persistent activity states for working memory, error-corrects, and integrates noisy cues. We consider the mechanisms by which simple and forgetful units can organize to collectively generate dynamics on the long time-scales required for such computations. We discuss the myriad potential uses of attractor dynamics for computation in the brain, and showcase notable examples of brain systems in which inherently low-dimensional continuous attractor dynamics have been concretely and rigorously identified. Thus, it is now possible to conclusively state that the brain constructs and uses such systems for computation. Finally, we look ahead by highlighting recent theoretical advances in understanding how the fundamental tradeoffs between robustness and capacity and between structure and flexibility can be overcome by reusing and recombining the same set of modular attractors for multiple functions, so they together produce representations that are structurally constrained and robust but exhibit high capacity and are flexible.

Introduction

The review examines how attractor dynamics may let simple neural elements generate stable representations, long timescales, memory, integration, and robustness. It traces attractor models from associative memory to evidence that brain circuits use them for essential computations.

Introduction: Attractor dynamics are proposed as a mechanism by which collective interactions among simple brain elements construct representations and support memory and integration.The review evaluates this hypothesis alongside predictions and experimental tests.
Introduction: Early circuit models addressed how neural circuits could generate distributed, stable activity patterns serving as associative memories.Hopfield networks stored multiple stable states learned from distributed input patterns.
Introduction: Continuous attractor networks were proposed to represent continuous variables, extending attractor models beyond discrete associative memories.The same principle was later applied to motor control, sensory amplification, memory, motion integration, decision making, and navigation.
Introduction: Model-derived predictions about connectivity and population correlations, combined with high-resolution recordings and rigorous analyses, provide extensive evidence for attractor networks in the brain.The review states that these networks support several essential computations.

What are attractors?

An attractor is a minimal set of states toward which nearby states flow, but applying this definition to neural circuits requires effective autonomy, state simplification, and treatment of noise. Attractors may be fixed points, manifolds, cycles, or chaotic sets.

What are attractors?: An attractor is the minimal set of states in a state space to which all nearby states eventually flow.A stable fixed point is the simplest example because neighboring states converge to it.
What are attractors?: Neural-circuit analysis often assumes effectively autonomous dynamics over timescales when inputs are constant and untuned across putative attractor states.The brain and its circuits cannot be completely isolated from external inputs.
What are attractors?: On timescales of seconds, neural states are commonly simplified to spike outputs or time-varying firing rates when those variables sufficiently predict circuit evolution.Parameters such as synaptic weights may be treated as fixed on short timescales but variable over longer ones.
What are attractors?: Attractors can be single states or manifolds supporting stationary, periodic, or chaotic trajectories.Continuous attractor manifolds may form curved and topologically complex shapes such as rings or tori.
What are attractors?: Noise displaces neural states from ideal attractors and can eventually cause escape, but nearby states still tend to converge and remain localized for extended periods.This noisy localization makes attractor states observable in autonomous dynamics.

Mechanisms: The construction of neural attractors

Neural attractors arise through recurrent feedback, symmetry, pattern formation, or finely tuned linear feedback, while inverse-problem analyses connect desired attractor states to generating network structures. Different mechanisms produce discrete, continuous, cyclic, or chaotic dynamics.

Mechanisms: The construction of neural attractors: Strong recurrent positive feedback counteracts activity decay and stabilizes selected neural states as attractors.The network’s shaping of positive feedback determines which states become stabilized.
Mechanisms: The construction of neural attractors: Forward and inverse analyses characterize how synaptic weights generate attractors and what network structures could produce specified attractor states.Lyapunov-function methods apply to symmetric weight matrices and rate-based dynamics, while inverse analysis is important because neural activity is easier to observe than synaptic weights.
Mechanisms: The construction of neural attractors: Hopfield networks create user-defined discrete attractors by Hebbian-like encoding of distributed activity patterns into symmetric weights.Partial or corrupted patterns can retrieve stored states, yielding content-addressable memory.
Mechanisms: The construction of neural attractors: Continuous symmetries in weights can generate continuous attractor sets, while local excitation and broader inhibition produce stable spatial patterns through pattern formation.These mechanisms support continuous manifolds rather than only isolated fixed points.
Mechanisms: The construction of neural attractors: Linear or near-linear neurons can support line, plane, or hyperplane attractors when recurrent feedback precisely cancels activity decay.Unlike pattern-forming systems, these long-lived attractors require finely tuned feedback strength.
Mechanisms: The construction of neural attractors: Strong asymmetric recurrent networks generically produce limit-cycle or chaotic attractors, which can remain much lower-dimensional than the number of neurons.Inhibition-dominated nonsymmetric networks instead have a zero-activity attractor with potentially large transient responses.

The potential utility of attractors for computation in the brain

Low-dimensional attractor dynamics support robust representation, memory, denoising, integration, sequence generation, evidence accumulation, and decision making, while exposing tradeoffs between robustness and drift.

Attractor networks provide stable internal states for reproducibly representing discrete or analog variables.
Low-dimensional attractors denoise corrupted inputs by correcting noise orthogonal to the attractor manifold.For a K-dimensional manifold in an N-neuron circuit, the corrected fraction is associated with K/N << 1.
Recurrent attractor dynamics are indispensable for persistent activity states and integration, whereas content-addressable memory and error reduction can also use feedforward computations.
Integration and decision making: Continuous attractors integrate signals over 1-100 seconds, far beyond the 10-100 millisecond integration timescale of single neurons.
Integration and decision making: Evidence accumulation and decision making can combine continuous integration with threshold-triggered transitions to discrete attractors.Winner-take-all networks implement this hybrid analog-discrete computation and can make noisy choices among N alternatives in ~log(N) time.
Sequence generation: Attractor dynamics generate robust sequences through low-dimensional limit cycles, but uncorrected noise along the attractor accumulates as timing variability.

Criteria for establishing attractor dynamics

Attractor dynamics are identified through low-dimensional population states, rapid recovery from perturbations, and invariance across diverse conditions rather than low dimensionality alone.

Attractor models predict states localized near a low-dimensional attractor set and rapid return after perturbation.
Putative attractor networks should be tested under conditions minimizing time-varying or localized external cues, while localization of the generating circuit is a separate question.
Simultaneous recordings of thousands of neurons enable direct characterization of whole-circuit low-dimensional state-space dynamics.
Low-dimensional embeddings can directly visualize manifolds below three dimensions and sometimes topologically simple higher-dimensional manifolds.
Low dimensionality and perturbation recovery are necessary but insufficient because inputs or feedforward projections can produce both signatures.
The decisive signature is invariance of population states and cell-cell relationships across time, rich input conditions, tuned-input removal, waking, and sleep.
The review reports that rigorous population analyses have established low-dimensional attractor dynamics across brain hierarchy levels and species.

Discrete attractors

Discrete attractors appear in cortical bistability and trained premotor dynamics, while broader discrete multistability remains less directly established than continuous attractor dynamics.

Bistability: Cortical up and down states switch between hyperpolarized and depolarized activity levels, persisting for hundreds of milliseconds to seconds.
Bistability: Up and down states can occur without thalamic or striatal input but tend to synchronize across cortex and striatum, suggesting a distributed origin.
Bistability: In a cued two-alternative delayed-response task, anterior lateral motor cortex neurons evolve during an approximately 1-second delay toward one of two response-guiding states.
Bistability: Training tailored ALM dynamics to a two-choice task suggests that slow plasticity may shape malleable recurrent structure in adult animals.
Discrete multistability: Candidate discrete multistable circuits include hippocampal, auditory, olfactory, and other systems showing global inhibition and selective recurrent excitation.
Discrete multistability: Direct evidence for discrete multistability is less exhaustive than evidence for continuous attractor networks, and quantitative tests of stability and invariance remain needed.

Continuous attractors

Continuous attractor dynamics support persistent, low-dimensional representations that integrate movement and sensory cues while remaining stable under perturbation and across conditions. Evidence from oculomotor, head-direction, grid-cell, and working-memory systems identifies these dynamics as concrete neural computations.

Oculomotor integrator: The oculomotor integrator is a line attractor that converts saccadic pulses into graded, persistent muscle-tension commands for stable gaze fixation.It also integrates smooth head-velocity signals, and network-level persistence does not arise from single neurons alone.
Head-direction cells: The mammalian head-direction circuit maintains heading by integrating rotational velocity and external cues on a one-dimensional ring manifold.Its states match measured head direction, and the ring can be re-anchored or reset by tuned external cues.
Head-direction cells: Natural perturbations in the head-direction circuit flow back to the ring, which remains invariant across waking and REM sleep.These findings validate fundamental predictions of continuous attractor and integrator models.
Grid cells: Co-modular grid cells form a two-dimensional state set whose internal relationships remain stable across environments, time, environmental rescaling, and sleep.Direct population recordings confirmed the predicted toroidal state-space structure and autonomous generation of these states.
Grid cells: Grid-cell population structure is preserved despite changes in spatial tuning, indicating that external inputs can alter how invariant internal states map onto cues.The proposed drivers include altered velocity estimation and feedforward inputs that shift network phase.
Shared mechanism: The head-direction and grid-cell systems support a shared pattern-formation principle involving local excitation or disinhibition combined with broader inhibition.This principle generates stationary continuous attractor states for neural computation and representation.
Graded working memory: In trained monkeys, a one-dimensional working-memory activity bump diffuses during the delay while retaining its profile, and its movement predicts later behavioral errors.The result is consistent with continuous-attractor models in which noise-driven perturbations affect remembered states.
Graded working memory: Unlike grid and head-direction networks, the working-memory attractor appears tailored to a trained multi-cue task and may have formed through learning.Its neural correlation structure might therefore change after training on other tasks.

Attracting limit cycles and trajectories

Attractor dynamics also describe periodic neural activity, but fixed-amplitude oscillations must be distinguished from merely transient or externally driven rhythms. Central pattern generators provide well-characterized examples of robust rhythmic attractors in motor systems.

Limit cycles: An attractive limit-cycle oscillator has an intrinsic, invariant amplitude, unlike a linear oscillator whose amplitude depends on its initial condition.Oscillations that decay, grow, or change long-term amplitude or frequency after perturbation are not limit cycles.
Limit cycles: Driven systems can display limit cycles because of their inputs rather than because of intrinsic attractor dynamics.The distinction matters when attributing observed oscillations to autonomous neural circuitry.
Central pattern generators: Central pattern generators are especially well-characterized candidate attractors because they robustly drive swimming, crawling, walking, breathing, and digestion.Their specific implementations differ across species while sharing common mechanisms and operating principles.
Related work: The review directs readers to existing literature for the broader study of periodic neural dynamics and related attractor systems.

Departures from low-dimensional continuous attractor dynamics

Evidence for low-dimensional continuous attractor dynamics is not uniform across neural circuits. V1, place-cell, and motor-cortical findings satisfy some attractor-related criteria, but important alternative explanations and unresolved tests remain.

Scope of evidence: The review emphasizes that some proposed low-dimensional attractor circuits either fail further testing or lack sufficient evidence to establish such dynamics.
V1: V1 orientation responses and sleep correlations satisfy some ring-attractor properties, but illusory-edge responses occur later than real-edge responses.The passage presents these observations as suggestive rather than definitive evidence for self-generated attractor dynamics.
Place cells: Place-cell representations are low-dimensional within individual environments, but multiple high-resolution maps exceed the capacity of a homogeneous attractor network.Place-cell correlations also remap across environments and are not preserved across long sleep bouts.
Place cells: Place-cell states may combine recurrent and feedforward dynamics, including inputs from grid cells, borders, landmarks, and reward sites.More detailed experimentation is needed before resolving whether autonomous low-dimensional dynamics explain the complex place-cell response.
Motor cortex: Stable low-dimensional trajectories in motor cortex have motivated limit-cycle and attractor hypotheses for movement generation, but stereotyped recording behaviors complicate interpretation.

Flexibility despite rigidity: modern glimpses into the broader potential of attractor networks

Attractor networks face a flexibility challenge because their robust, low-dimensional states are rigidly structured. The review describes modular and integrator-based strategies that reuse attractors, expand capacity, and rapidly construct representations for new variables.

Rigid, low-dimensional attractor states support robust computation but appear to conflict with the flexibility required for representation, memory, and computation.
Integrator-based mapping replaces exhaustive state-by-state associations with one anchor and a learned velocity-to-shift projection.The same recurrent attractor circuit can generate representations for unvisited external states through displacement.
A single attractor can be reused for multiple variables by adding anchor points and coupling each variable’s velocity cues to the shared shift mechanism.This permits switching between variables without reconfiguring the recurrent attractor.
Fully connected symmetric networks scale to approximately N chosen attractor states, while D-dimensional resolution P per dimension requires N ≈ P^D neurons.This capacity scaling creates a curse of dimensionality that motivates modular constructions.
Modular subnetworks expand capacity by combining independently updateable attractors, yielding approximately N^M combinations from M modules with approximately N states each.Coupled modules can preserve denoising while producing structured combinations rather than arbitrary Hopfield states.
Mixed modular integrators represent input spaces of dimension D ≤ MK by reusing M modules of dimension K, while excess capacity trades resolution against dimensionality.Random projections distribute each external dimension across the modules, enabling flexible reuse without recurrent rewiring.

Looking ahead

The review concludes that attractor theory now unifies major computational functions and is supported by direct population-level evidence in brain circuits. It also identifies open mechanistic questions and opportunities to study attractor interactions and reuse.

Attractor theory provides a unifying framework for integration, representation, memory, error-correction, and efficient learning and inference.Population-wide physiology has enabled direct visualizations of attractor dynamics in the brain.
Artificial neural networks trained on memory, integration, and decision tasks develop attractor dynamics across navigation, vision, and language.Preconfigured attractors may support faster, more data-efficient, and more generalizable learning.
Open questions include whether low-firing-rate synchronous spiking networks can support attractors while combining synchronization and oscillatory phase dynamics.
Continuous-attractor development models range from unsupervised associative plasticity and error-driven rules to backpropagation with architectural constraints.
Future experiments will use high-resolution population recordings and perturbations across brain areas, while theory will examine the formation, interaction, and reuse of multiple low-dimensional structures.
Figure 1 relates connectivity motifs to stable population patterns and state-space dynamics, including Hopfield, winner-take-all, ring, toroidal, line, and limit-cycle attractors.
Figure 2 maps attractor properties onto representation, denoising, classification, integration, decision-making, reuse, high-capacity modular networks, and mixed modular coding.
Figures 3–6 provide brain evidence for discrete, linear, ring, and toroidal attractor dynamics in cortical, premotor, oculomotor, head-direction, and grid-cell systems.