CANNs: Continuous Attractor Neural Networks Toolkit

中文说明请见 README_zh.md

Overview

CANNs (Continuous Attractor Neural Networks toolkit) is a research toolkit built on BrainPy and JAX, with optional Rust-accelerated canns-lib for selected performance-critical routines (e.g., TDA/Ripser and task generation). It bundles model collections, task generators, analyzers, and the ASA pipeline (GUI/TUI) so researchers can run simulations and analyze results in a consistent workflow. The API separates models, tasks, analyzers, and trainers to keep experiments modular and extensible.

Architecture

Layer hierarchy of the CANNs library showing five levels: Application (Pipeline orchestration), Functional (Task, Trainer, Analyzer, Utils modules), Core Models (CANN implementations), Foundation (BrainPy/JAX and Rust FFI backends), and Hardware (CPU/GPU/TPU support)

The CANNs library follows a modular architecture guided by two core principles: separation of concerns and extensibility through base classes. The design separates functional responsibilities into five independent modules:

1. Models (canns.models) define neural network dynamics;
2. Tasks (canns.task) generate experimental paradigms and input data;
3. Analyzers (canns.analyzer) provide visualization and analysis tools;
4. Trainers (canns.trainer) implement learning rules for brain-inspired models;
5. Pipeline (canns.pipeline) orchestrates complete workflows.

Each module focuses on a single responsibility—models don't generate input data, tasks don't analyze results, and analyzers don't modify parameters. This separation ensures maintainability, testability, and extensibility. All major components inherit from abstract base classes (BasicModel, BrainInspiredModel, Trainer) that define standard interfaces, enabling users to create custom implementations that seamlessly integrate with the built-in ecosystem.

Core Features

• Model collections: basic CANNs (1D/2D, SFA), hierarchical path integration, theta-sweep models, brain-inspired models (e.g., Amari-Hopfield, linear/spiking layers)
• Task generators: smooth tracking, population coding, template matching, open/closed-loop navigation
• Analyzer suite: energy landscapes, tuning curves, raster/firing-rate plots, TDA and decoding utilities, cell classification
• ASA pipeline & GUI/TUI: end-to-end workflow for preprocessing, TDA, decoding, and result visualization (e.g., CohoMap/CohoSpace/PathCompare/FR/FRM/GridScore)
• Training & extensibility: HebbianTrainer plus base classes for consistent extension
• Optional acceleration: canns-lib for selected performance-critical routines

Analyzer Visuals

Overview of Neural Dynamics Models. Comparison of three basic models: (A) 1D CANN, (B) 2D CANN, and (C) Grid Cell Network

Rich Analyzer Visualization Results

ASA GUI preview

ASA GUI demo video

Smooth Tracking 1D Activity bump following a moving stimulus	CANN2D Encoding 2D population encoding patterns over time
Theta Sweep Animation Theta-modulated sweep dynamics
Bump Analysis Bump fitting and stability diagnostics	Torus Bump Bump dynamics projected onto a torus manifold

Quick Start

1D CANN smooth tracking (imports → simulation → visualization)

import brainpy.math as bm
from canns.analyzer.visualization import PlotConfigs, energy_landscape_1d_animation
from canns.models.basic import CANN1D
from canns.task.tracking import SmoothTracking1D

# simulation time step
bm.set_dt(0.1)

# build model
cann = CANN1D(num=512)

# build tracking task (Iext length = duration length + 1)
task = SmoothTracking1D(
    cann_instance=cann,
    Iext=(0.0, 0.5, 1.0, 1.5),
    duration=(5.0, 5.0, 5.0),
    time_step=bm.get_dt(),
)
task.get_data()


# one-step simulation callback
def step(t, stimulus):
    cann(stimulus)
    return cann.u.value, cann.inp.value


# run simulation loop
us, inputs = bm.for_loop(
    step,
    operands=(task.run_steps, task.data),
)

# visualize with energy landscape animation
config = PlotConfigs.energy_landscape_1d_animation(
    time_steps_per_second=int(1 / bm.get_dt()),
    fps=20,
    title="Smooth Tracking 1D",
    xlabel="State",
    ylabel="Activity",
    show=True,
)

energy_landscape_1d_animation(
    data_sets={"u": (cann.x, us), "Iext": (cann.x, inputs)},
    config=config,
)

Installation

# CPU-only
pip install canns

# Optional accelerators (Linux)
pip install canns[cuda12]
pip install canns[cuda13]
pip install canns[tpu]

# GUI (ASA Pipeline)
pip install canns[gui]

Optional (uv):

uv pip install canns

Docs & Examples

• Documentation and tutorials: https://routhleck.com/canns/
• Local scripts: examples/
• Sphinx docs and notebooks: docs/
• ASA GUI entry: canns-gui

Citation

If you use CANNs in your research, please cite:

@software{he_2026_canns,
  author       = {He, Sichao and
                  Tuerhong, Aiersi and
                  She, Shangjun and
                  Chu, Tianhao and
                  Wu, Yuling and
                  Zuo, Junfeng and
                  Wu, Si},
  title        = {CANNs: Continuous Attractor Neural Networks Toolkit},
  month        = feb,
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v1.0.0},
  doi          = {10.5281/zenodo.18453893},
  url          = {https://doi.org/10.5281/zenodo.18453893}
}

Plain text:

He, S., Tuerhong, A., She, S., Chu, T., Wu, Y., Zuo, J., & Wu, S. (2026). CANNs: Continuous Attractor Neural Networks Toolkit (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.18453893

Contributing & License

Contributions are welcome. Please read CONTRIBUTING.md before opening a PR.

Apache License 2.0. See LICENSE.