canns-lib

High-performance computational acceleration library for CANNs (Continuous Attractor Neural Networks), providing optimized Rust implementations for computationally intensive tasks in neuroscience and topological data analysis.

Overview

canns-lib is a modular library designed to provide high-performance computational backends for the CANNS Python package. It currently includes the Ripser module for topological data analysis, with plans for additional modules covering approximate nearest neighbors, dynamics computation, and other performance-critical operations.

Modules

🔬 Ripser - Topological Data Analysis

High-performance implementation of the Ripser algorithm for computing Vietoris-Rips persistence barcodes. Drop-in replacement for ripser.py with identical output (verified at the level of bar counts and per- dimension birth/death values on dense and sparse inputs).

Performance vs ripser.py (v0.9.0)

Measured by benchmarks/ripser/harness/comprehensive_benchmark.py --fast (24 dense point-cloud tests, n ∈ {50, 100, 150, 300}, maxdim ∈ {1, 2}, categories: circle, sphere, torus, random, clusters, grid, swiss_roll, moons, circles). Cross-validated on macOS arm64 and Linux x86_64 (16-core A100 server, RAYON_NUM_THREADS=16).

Headline Linux maxdim=2 result: peak 1.91× on torus n=300. Per-dimension persistence values match ripser.py exactly on both platforms (counts_match=True | birth/death values match=True).

Performance vs ripser.py (v0.8.0 baseline, before this release)

Same harness, same matrices:

Shuffle null models with explicit parameters

canns_lib.ripser.shuffle_null_model independently circular-shifts each feature column, computes distances between rows, and runs Ripser. Its metric and persistence options are explicit:

from canns_lib.ripser import ripser, shuffle_null_model

# X has shape (timepoints, features).
real = ripser(X, metric="cosine", maxdim=2, coeff=47)
null = shuffle_null_model(
    X,
    num_shuffles=100,
    metric="cosine",
    maxdim=2,
    coeff=47,
    seed=17,
)

This generic feature-space null does not perform the complete ASA analysis. The companion CANNs PR #103 keeps activity selection, standardization, PCA, density selection and graph construction in CANNs, configured through TDAConfig and rerun each round. Do not substitute a shuffle of an already processed point cloud for that full workflow. Independent column shifts also do not define a valid null for a precomputed distance matrix, so distance_matrix=True is rejected.

Explicit shifts support exact replay; generate_offsets exposes the same offset generation for application workflows. max_workers=1 is the default, with bounded thread concurrency available. Failed rounds raise with replay information instead of contributing zero or disappearing from the null.

The default result is a per-dimension list of maximum finite lifetimes. return_details=True also returns shifts, essential-class counts and complete diagrams. Finite maxima alone do not test essential classes.

The previous private _ripser_core.shuffle_null_model computed a different neuron-distance null and has been removed; calls now raise a migration error. Its timing ratios against the full ASA workflow were not measurements of the same analysis and are no longer presented as acceleration results. Downstream ASA callers must use the coordinated CANNs update.

The public Rust fuzzy_union kernel can be used within that pipeline to build an owned dense float64 adjacency matrix without fixing its scientific parameters. See the shuffle API, migration and kernel guide for contracts, memory costs and examples. The generic shuffle API alone does not establish agreement with the ASA workflow or any paper.

Features

• Algorithmic improvements: Row-by-row edge generation, cleared coboundary enumeration
• Memory optimization: Structure-of-Arrays reduction matrix, k-major binomial coefficient table, packed (index, coefficient) EntryT (24 → 16 bytes), GAT-based static-dispatch cofacet enumeration with inline simplex-vertex stack array
• Parallel processing: Multi-threading with Rayon (enabled by default)
• Full Compatibility: Drop-in replacement for ripser.py with identical API
• Multiple Metrics: Support for Euclidean, Manhattan, Cosine, and custom distance metrics
• Sparse Matrices: Efficient handling of sparse distance matrices
• Cocycle Computation: Optional computation of representative cocycles
• Configurable feature-space shuffle: explicit metric and persistence parameters, reproducible per-feature shifts, bounded concurrency and explicit failures; optional Rust fuzzy_union for dense adjacency construction (API guide)

Two experimental paths are kept under env-flag opt-in only: CANNS_RIPSER_USE_LOCKFREE=1 and CANNS_RIPSER_APPARENT=1. Both are correctness-fixed and match ripser.py outputs but are currently net- slower than sequential on this codebase.

🧭 Spatial Navigation (RatInABox parity)

Accelerated reimplementation of RatInABox environments and agents with PyO3/ Rust. Supports solid and periodic boundaries, arbitrary polygons, holes, and thigmotaxis wall-following.

Performance Snapshot

The spatial backend delivers ~700× runtime speedups vs. the pure-Python reference when integrating long trajectories. Benchmarked with benchmarks/spatial/step_scaling_benchmark.py (dt=0.02, repeats=1).

Plots and CSV summaries are emitted to benchmarks/spatial/outputs/.

Highlights

• Full parity with RatInABox API (Environment, Agent, trajectory import/export)
• Polygon & hole support with adaptive projection and wall vectors
• Parity comparison tools in example/trajectory_comparison.py
• Visualization utilities: drop-in replacements for RatInABox's plotting helpers (trajectory, heatmaps, histograms)
• Benchmark scripts for long-step drift and speedup under benchmarks/spatial/

Visualization Helpers

from canns_lib import spatial

env = spatial.Environment(dimensionality="2D", boundary_conditions="solid")
agent = spatial.Agent(env, rng_seed=2025)

for _ in range(2_000):
    agent.update(dt=0.02)

# Trajectory with RatInABox-style colour fading and agent marker
agent.plot_trajectory(color="changing", colorbar=True)

# Other helpers mirror RatInABox naming
agent.plot_position_heatmap()
agent.plot_histogram_of_speeds()
agent.plot_histogram_of_rotational_velocities()

See example/spatial_plotting_demo.py for a full script that produces the trajectory, heatmap, and histogram figures showcased above.

Drift Velocity Control

Guide agent movement toward target directions while maintaining natural stochastic motion (matches RatInABox API):

# Basic drift usage
agent.update(
    dt=0.02,
    drift_velocity=[0.05, 0.02],  # Target velocity vector
    drift_to_random_strength_ratio=5.0  # Drift strength relative to random motion
)

Parameters:

• drift_velocity: Target velocity vector (must match environment dimensionality)
• drift_to_random_strength_ratio: Controls balance between drift and randomness
  • 0.0 = pure random motion (no drift)
  • 1.0 = equal weighting (default)
  • > 1.0 = stronger drift toward target

Use cases: Goal-directed navigation, reinforcement learning, biased exploration.

See example/drift_velocity_demo.py for detailed examples with visualizations.

🚀 Coming Soon

• Dynamics: High-performance dynamics computation for neural networks
• And more...

Installation

From PyPI (Recommended)

pip install canns-lib

From Source

git clone https://github.com/Routhleck/canns-lib.git
cd canns-lib
pip install maturin
maturin develop --release

Quick Start

Using the Ripser Module

import numpy as np
from canns_lib.ripser import ripser

# Generate sample data
data = np.random.rand(100, 3)

# Compute persistence diagrams
result = ripser(data, maxdim=2)
diagrams = result['dgms']

print(f"H0: {len(diagrams[0])} features")
print(f"H1: {len(diagrams[1])} features")
print(f"H2: {len(diagrams[2])} features")

Advanced Options

# High-performance computation with progress tracking
result = ripser(
    data,
    maxdim=2,
    thresh=1.0,                    # Distance threshold
    coeff=2,                       # Coefficient field Z/2Z
    do_cocycles=True,              # Compute representative cycles
    verbose=True,                  # Detailed output
    progress_bar=True,             # Show progress
    progress_update_interval=1.0   # Update every second
)

# Access results
diagrams = result['dgms']          # Persistence diagrams
cocycles = result['cocycles']      # Representative cocycles
num_edges = result['num_edges']    # Number of edges in complex

Sparse Matrix Support

from scipy import sparse

# Create sparse distance matrix
row = [0, 1, 2]
col = [1, 2, 0]
data = [1.0, 1.5, 2.0]
sparse_dm = sparse.coo_matrix((data, (row, col)), shape=(3, 3))

# Compute with sparse matrix (automatically detected)
result = ripser(sparse_dm, distance_matrix=True, maxdim=1)

Compatibility

The ripser module maintains 100% API compatibility with ripser.py:

# These work identically
import ripser as original_ripser
from canns_lib.ripser import ripser

result1 = original_ripser.ripser(data, maxdim=2)
result2 = ripser(data, maxdim=2)

# Results are numerically identical
assert np.allclose(result1['dgms'][0], result2['dgms'][0])

Development

Building from Source

# Prerequisites
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
pip install maturin

# Build and install
git clone https://github.com/Routhleck/canns-lib.git
cd canns-lib
maturin develop --release --features parallel

# Run tests
python -m pytest tests/ -v

Optional build features

All of these are off by default, so the released wheel and pip install -e . are unchanged unless you ask for them.

mimalloc (opt-in, and off by default because mimalloc 0.1.43 segfaults on macOS Darwin 26) and the timing/counter probes (apparent-stats, gpu-phase-stats, walk-probe) are documented in Cargo.toml, where each one says what it measures and what it costs.

The published wheels are CPU-only

The gpu feature is not in the release build. [tool.maturin] sets features = ["parallel", "hugepage", "extension-module"], so a wheel installed from PyPI contains no CUDA code at all: the apparent-pair pre-pass runs on the CPU, and the barcode is identical either way. The GPU pre-pass is an opt-in build, not a download.

To get it, build the extension yourself:

# keep `extension-module` in the list: it is what makes PyO3 produce a
# cdylib for Python instead of a binary linked against libpython
maturin develop --release --features parallel,hugepage,extension-module,gpu

A CUDA toolkit is not required to build it. build.rs compiles kernel/ripser_kernels.cu with nvcc into a cubin and embeds it when a toolkit is on PATH; without one the build still succeeds and the driver JIT-compiles the same kernel file through NVRTC at run time. If there is no usable device, or the driver refuses the kernel, the run falls back to the CPU pre-pass — the library never requires a GPU to be correct, and the barcode does not change.

To see which path a run actually took, set CANNS_RIPSER_STATS=1 and read the gpu_err field of the STATS line on stderr: empty means the GPU pre-pass ran, and anything else is the reason it did not. A silent CPU fallback is otherwise indistinguishable from "the GPU was never tried".

CANNS_RIPSER_STATS=1 python -c "
from canns_lib.ripser import ripser
ripser(data, maxdim=2)
" 2>&1 | grep -o 'gpu_err=\"[^\"]*\"'

The library uses device 0. CUDA_VISIBLE_DEVICES remaps which physical device that is, which is what you want on a host where another job holds the other device:

CUDA_VISIBLE_DEVICES=1 python -c "from canns_lib.ripser import ripser; ..."

The benchmark harness in benchmarks/ripser/ has the same knob under the name CANNS_BENCH_GPU (it sets CUDA_VISIBLE_DEVICES for the child process), and it records the device in every row of every table it writes.

Running Benchmarks

cd benchmarks
python compare_ripser.py --n-points 100 --maxdim 2 --trials 5

benchmarks/ripser/ also carries the head-to-head harness against ripser++, the standard-suite runner, the scale scan and the load characterisation of the grid-cell corpus. Each table is produced by a runner, and every row records the number of cores used by processes that are not ours while that row was measured, so a table measured on a busy host can be told apart from one that was not. The tables themselves are not in the repository -- the grid-cell corpus they draw on has an unsettled licence -- and they are kept in the experiment record next to the analysis that turns them into numbers.

Technical Details

Ripser Module Architecture

• Dual API paths: High-performance versions and full-featured versions with progress tracking
• Memory optimization: Structure-of-Arrays layout, intelligent buffer reuse
• Sparse matrix support: Efficient handling via neighbor intersection algorithms
• Progress tracking: Built-in progress bars using tqdm when available
• Parallel processing: Multi-threading with Rayon

Algorithmic Optimizations

• Dense edge enumeration: O(n²) row-by-row generation vs O(n³) vertex decoding
• Sparse queries: O(log k) binary search vs O(k) linear scan
• Cache-friendly data structures: SoA matrix layout, k-major binomial tables
• Packed simplex entries: (index, coefficient) packed into one 64-bit word, halving DiameterEntryT size from 24 → 16 bytes

License

Licensed under the Apache License, Version 2.0. See LICENSE for details.

Citation

If you use canns-lib in your research, please cite the toolkit paper (which describes both canns and canns-lib):

@misc{he2026canns,
  title        = {CANNs: A Toolkit for Research on Continuous Attractor Neural Networks},
  author       = {He, Sichao and
                  Tuerhong, Aiersi and
                  She, Shangjun and
                  Chu, Tianhao and
                  Wu, Yuling and
                  Zuo, Junfeng and
                  Wu, Si},
  year         = 2026,
  eprint       = {2606.27783},
  archivePrefix = {arXiv},
  primaryClass  = {q-bio.NC},
  doi          = {10.48550/arXiv.2606.27783},
  url          = {https://arxiv.org/abs/2606.27783}
}

Plain text:

He, S., Tuerhong, A., She, S., Chu, T., Wu, Y., Zuo, J., & Wu, S. (2026). CANNs: A Toolkit for Research on Continuous Attractor Neural Networks. arXiv:2606.27783. https://arxiv.org/abs/2606.27783

If you want to cite this Rust backend specifically, you can additionally reference the Zenodo archive:

@software{canns_lib,
  title={canns-lib: High-Performance Computational Acceleration Library for CANNS},
  author={He, Sichao},
  url={https://github.com/Routhleck/canns-lib},
  year={2025}
}

Acknowledgments

Ripser Module

• Ulrich Bauer: Original Ripser algorithm and C++ implementation
• Christopher Tralie & Nathaniel Saul: ripser.py Python implementation
• Rust community: Amazing ecosystem of high-performance libraries

Related Projects

• Ripser: Original C++ implementation
• ripser.py: Python bindings for Ripser
• CANNS: Continuous Attractor Neural Networks
• scikit-tda: Topological Data Analysis in Python