CANNs: Attractor Networks and Beyond

• CANNs are polysemous technical labels, primarily representing continuous attractor neural networks with localized bump states and translationally invariant connectivity.
• The 'canns' toolkit is an open-source framework that unifies simulation, analysis, and inference via integrated Python, Rust, and topological modules.
• CANNs also denote constitutive, actuarial, and continually-adaptive models, facilitating advanced applications in mechanics, insurance pricing, and continual learning.

CANNs, or canns in the software context, is a polysemous technical label in current research. Its canonical neuroscience meaning is the continuous attractor neural network, a recurrent population model with translationally invariant connectivity that supports a continuum of localized bump states for variables such as spatial position, head direction, and movement direction. The same label also denotes a recent open-source toolkit for CANN research, and unrelated acronymic model families in solid mechanics, actuarial science, and lifelong learning, including constitutive artificial neural networks, combined actuarial neural networks, and continually-adaptive neural networks (He et al., 26 Jun 2026, Tacke et al., 22 May 2026, Holvoet et al., 2023, Shahawy et al., 2022).

1. Terminological scope

In recent arXiv usage, the label appears in several non-equivalent senses. This suggests that canns is best interpreted contextually rather than as a single unified concept.

Sense of CANNs/canns	Domain	Representative source
Continuous attractor neural networks	Theoretical and computational neuroscience	(He et al., 26 Jun 2026)
canns toolkit	CANN simulation, analysis, and experimental-data inference	(He et al., 26 Jun 2026)
Constitutive artificial neural networks	Hyperelastic constitutive modeling	(Tacke et al., 22 May 2026)
Combined actuarial neural networks	Insurance pricing	(Holvoet et al., 2023)
Continually-Adaptive Neural Networks	NAS and continual learning	(Shahawy et al., 2022)

The ambiguity is substantive rather than stylistic. In neuroscience, CANNs are attractor systems over continuous state spaces; in mechanics, they are energy-based constitutive models; in actuarial modeling, they are neural corrections on top of GLM or GBM baselines; and in lifelong learning, they are a proposed paradigm joining automation and continual adaptation (Tacke et al., 22 May 2026, Holvoet et al., 2023, Shahawy et al., 2022).

2. Continuous attractor neural networks as the canonical sense

The canonical CANN formalism is a recurrent neural field or discretized population model in which recurrent coupling depends on relative, rather than absolute, preferred stimulus values. In the toolkit formulation, the core equations are

$$W_{ij}=\frac{J_0}{\sqrt{2\pi}\,a}\exp\!\bigl(-d_{ij}^2/2a^2\bigr), \qquad \tau\,\dot u_i=-u_i+\sum_j W_{ij}r_j+I_i,$$

with divisive normalization

$r_i=\frac{[u_i]_+^2}{1+k\sum_j [u_j]_+^2}.$

This architecture produces localized bump states, translation invariance, contrast-invariant tuning, and stable attractor manifolds (He et al., 26 Jun 2026).

A closely related two-dimensional formulation used for visual tracking writes

$$\tau_c \frac{\partial U(x,t)}{\partial t} = -U(x,t) + I_{\mathrm{ext}}(x,t) + \rho \int_{\mathbb{R}^2} J(x-x')\, r(x',t)\, dx',$$

with Gaussian kernel

$$J(x-x')=\frac{A}{2\pi a^2}\exp\!\left(-\frac{\|x-x'\|^2}{2a^2}\right).$$

Because of translational invariance, the network admits a continuous family of stationary bump states

$$U(x\mid z)=U_0\exp\!\left(-\frac{\|x-z\|^2}{4a^2}\right), \qquad r(x\mid z)=r_0\exp\!\left(-\frac{\|x-z\|^2}{2a^2}\right),$$

where the bump center $z$ can be any point in the state space (Zhou et al., 21 Jun 2026).

Classical CANN intuition is summarized in later work as local excitation and global inhibition, approximately symmetric couplings, continuous attractor manifolds or slowly drifting stable activity “bumps,” and dynamics evolving on a low-dimensional manifold of near-equivalent states. That characterization is central to how later models borrow from CANNs without necessarily instantiating exact continuous attractors (Zhang et al., 13 Jun 2026).

3. Dynamical regimes, plasticity, and control

A large part of the CANN literature concerns how slow suppressive processes reshape bump dynamics. With short-term synaptic depression (STD) and short-term facilitation (STF), dynamic synapses alter mobility, persistence, and decoding properties. One study reports that STD produces slow-decaying plateau behaviors, increases mobility, enhances tracking of time-varying stimuli, and can yield anticipative or overtaking responses, whereas STF stabilizes responses to noisy inputs and improves decoding accuracy by retaining a short memory trace of recent stimulus history (Fung et al., 2011).

In two-dimensional CANNs, STD and spike-frequency adaptation (SFA) generate qualitatively similar phase structure. The reported phases are static, moving, bistable, and silent, and the perturbative framework predicts phase diagrams, dynamical variables, translational instability, and the intrinsic speed of spontaneous motion (Fung et al., 2015). A related analysis of CANNs with STD describes an even richer repertoire, including stationary bumps, moving bumps, uniform firing, population spikes, sloshers, emitter states, quasi-periodic oscillations, and chaos, organized by the interaction of recurrent excitation, divisive inhibition, and slow synaptic recovery (Wang et al., 2015).

The anticipative-tracking literature gives a hardware realization of these ideas. In a spintronics implementation, short-term depression is realized by a magnetic tunnel junction, and the resulting dynamic synapses are embedded in one-dimensional and two-dimensional CANNs. The reported outcome is prediction of moving objects via micromagnetic simulations, with no training, using the negative feedback induced by synaptic depression to shift the bump ahead of the stimulus (Zheng et al., 2020).

Recent machine-learning work often uses CANNs as an inductive bias rather than as a literal dynamical-system model. The Controlled Dynamics Attractor Transformer explicitly states that it is not a classical CANN, but a CANN-inspired control mechanism for attractor dynamics. Its continuous-time state equation adds passive decay, topology-conditioned mutual excitation, global damping, and semantic energy descent,

$$\tau \dot{\mathbf{x}} = W\mathbf{x} - (1+\omega)\mathbf{x} - \nabla_{\mathbf{g}}E,$$

and its kernelized coupling induces topology-aware discrete attractors rather than a true continuous manifold. The paper’s stated goal is to preserve the CANN flavor of structured excitation/inhibition and stable attractor evolution inside a modern energy-based Transformer (Zhang et al., 13 Jun 2026).

4. The canns toolkit

The lowercase form canns refers specifically to an open-source toolkit intended to unify the CANN research workflow. The toolkit combines three tightly integrated components: canns, a Python library on BrainPy/JAX; canns-lib, a Rust acceleration backend; and ASA (Attractor Structure Analyzer), a PySide6 pipeline applying persistent homology and cohomology to experimental neural recordings (He et al., 26 Jun 2026).

The Python stack is organized into five layers: Application / Pipeline, Functional, Core Models, Foundation, and Hardware. Its model layer includes standardized 1D and 2D CANNs, SFA variants, grid cell networks, theta-sweep models, hierarchical path-integration models, and brain-inspired recurrent or spiking models. The functional layer includes task generators, trainer modules, analyzer modules, and utilities; the trainer interface explicitly names HebbianTrainer, OjaTrainer, SangerTrainer, BCMTrainer, and STDPTrainer (He et al., 26 Jun 2026).

ASA formalizes attractor-geometry inference from neural recordings. It constructs time-indexed or spatial point clouds of population activity, computes Vietoris–Rips persistent homology, and uses persistent cohomology to recover circular coordinates. The ideal topological signatures are reported as

$$S^1:(\beta_0,\beta_1,\beta_2)=(1,1,0), \qquad T^2:(\beta_0,\beta_1,\beta_2)=(1,2,1),$$

corresponding to ring and torus attractors, respectively (He et al., 26 Jun 2026).

The toolkit is also a performance-engineering project. The Rust backend reports 1.13× average speedup and 1.82× peak speedup for Ripser-based persistent homology, and about 700× speedup for spatial navigation workloads, with the $10^6$-step benchmark reported as 0.27 s vs 192.8 s. The package ships with reproducible pipelines for SFA-driven anticipative tracking, theta sweeps in head-direction/place/grid systems, hierarchical path integration, and ASA analyses of medial entorhinal cortex recordings (He et al., 26 Jun 2026).

5. Self-organization, hybridization, and efficient replication

Recent work has moved beyond hand-specified attractor connectivity. A theoretical framework based on Hebbian plasticity, firing-rate adaptation, and global inhibition shows that translationally invariant inputs can drive the self-organization of Gaussian feedforward weights, yielding CANN-like structure as an equilibrium solution rather than a hard-wired prior. The same work states that anticipatory shifts can emerge in a purely feedforward architecture, that multilayer networks amplify this predictive shift approximately linearly, and that adding recurrent plasticity can produce self-sustaining moving bumps and speed-modulated unidirectional path integration (Emina et al., 12 Jun 2026).

A separate line of work develops population-scale ANN–CANN hybridization for visual object tracking. The Hybrid Tracking Neural Network aligns ANN response maps and CANN dynamics in the same target-state space, interprets ANN estimates as asymptotically unbiased but noise-sensitive, and CANN estimates as lower-variance but temporally lagged, and fuses the two representations before decoding. On nine benchmarks—OTB50, OTB100, GOT-10k, LaSOT, TColor128, UAV123, NfS, VOT2019, and TrackingNet—the reported result is the best precision and success rate on 8 of the 9 benchmarks, with GOT-10k as the exception (Zhou et al., 21 Jun 2026).

Efficiency-oriented navigation research takes a different direction: replacing recurrent CANN simulation with learned surrogates. In path integration for dead reckoning, lightweight ANN modules are trained to replicate the neurodynamic patterns of CANN-based head direction cells and grid cells. The reported outcome is NeuroSLAM-level positioning accuracy together with efficiency improvements of approximately 17.5% on the general-purpose device and 40~50% on the edge device; the detailed experiments additionally report 39.4% faster on Raspberry Pi 4B in one setting and 48.9% reduction in execution time in another (Ge et al., 21 Nov 2025). A plausible implication is that recent CANN research is increasingly bifurcating into two agendas: preserving attractor dynamics explicitly, and approximating them when deployment efficiency dominates.

6. Constitutive and actuarial meanings of CANNs

In solid mechanics, constitutive artificial neural networks are hybrid or grey-box constitutive models for hyperelastic materials. Rather than learning stress directly, a CANN learns a strain-energy density $\Psi$ as a function of deformation invariants and obtains stress by differentiation. In the isotropic incompressible setting,

$r_i=\frac{[u_i]_+^2}{1+k\sum_j [u_j]_+^2}.$0

The stated advantages are objectivity, compatibility with isotropic material symmetry, thermodynamic consistency, and practical generalization relative to direct stress regressors (Tacke et al., 22 May 2026).

Two recent LLM-centered papers use these constitutive CANNs as the concrete substrate for automated model generation. One introduces a Creator–Inspector framework in which a Creator proposes a CANN and an Inspector audits it against nine physical constraints: thermodynamic consistency, stress symmetry, objectivity, material symmetry, polyconvexity / ellipticity, growth condition, energy normalization, stress normalization, and non-negativity of strain energy. Across 10 runs on 3 benchmark problems, adding the Inspector raises the share of exported models satisfying all physical constraints from 91% to 100% for Claude Opus 4.7 and from 37% to 56% for Kimi K2.5 (Tacke et al., 22 May 2026). The other paper presents GenCANN, an OpenAI O3-driven framework that generates a CANN from a material class and dataset, performs three refinement rounds, and reports a stability analysis over 124 individual implementations, of which about 31% had syntax errors and 5% had training errors on first try before retries fixed them (Tacke et al., 1 Dec 2025).

In insurance pricing, combined actuarial neural networks are neither attractor systems nor constitutive models. They start from an actuarial baseline prediction—GLM in the original formulation and GBM in the paper’s extension—and learn a neural-network correction on top of that baseline. The paper states that CANN GBM flexible is usually the best or among the best frequency model, that CANNs often have the best calibration overall relative to plain FFNNs, and that their learned structure can be distilled into a global surrogate GLM to obtain a technical tariff table (Holvoet et al., 2023).

7. Continually-Adaptive Neural Networks

In the NAS–CL literature, CANNs denotes Continually-Adaptive Neural Networks, a proposed paradigm at the intersection of Neural Architecture Search and Continual Learning. The review defines an ANN as continually-adaptive if it is fully-autonomous throughout its life-cycle, inherently continual, and capable of handling an infinite stream of incremental input while allowing some performance degradation over time. The intended system is not a static model retrained repeatedly, but one that can choose or modify its architecture, learn new tasks continually, retain prior knowledge, and adapt without manual redesign (Shahawy et al., 2022).

The same review formalizes the CANN paradigm in terms of Data automation and Model plasticity, and lists desiderata including domain-agnostic behavior, abstract representation, dynamic architecture with bounded capacity, graceful forgetting, few-shot learning capability or positive forward transfer, multi-modal input support, architecture search space inference, no explicit data storage for previous tasks, and learning-rule plasticity. It also emphasizes unresolved problems: computational cost, lack of standardized evaluation, search-space dependence, bounded memory under unbounded tasks, dynamic task-boundary inference, and safety under continuous autonomous data collection (Shahawy et al., 2022).

This broader usage is conceptually disconnected from continuous attractor neural networks, but it illustrates the degree to which the acronym has been repurposed across disciplines. The shared orthography therefore does not imply shared mathematics, shared objectives, or shared implementation practice; it marks a family of domain-specific research programs whose overlap is largely terminological rather than substantive.