JUNIPER

GPU Accelerated Python Implementation of CEDAR

JUNIPER is a high-performance framework for simulating neural dynamics architectures on the GPU. Built on top of JAX, it provides a flexible Python API for constructing computational graphs of processing steps, compiling them, and running simulations with automatic GPU acceleration.

What is JUNIPER?

JUNIPER implements the concepts of the CEDAR framework in Python, leveraging JAX for just-in-time (JIT) compilation and GPU execution. It is designed for researchers and engineers working with Dynamic Field Theory (DFT) and neural dynamics, but its modular step-based architecture makes it suitable for any signal-processing pipeline that benefits from GPU acceleration.

Key Features

• GPU acceleration via JAX with automatic JIT compilation
• Modular step-based architecture -- compose complex processing pipelines from reusable building blocks
• Pythonic connection syntax using >> and << operators to wire steps together
• Rich library of built-in steps covering algebra, array manipulation, neural fields, image processing, robotics, and I/O
• Dynamic and static steps -- efficient separation of time-varying (dynamic) and feed-forward (static) computation
• Buffer save/restore for persisting simulation state across runs

Quick Start

from juniper import GaussInput, NeuralField, Gaussian, StaticGain
from juniper.Architecture import get_arch

# Create steps
gi = GaussInput("input", {"shape": (50,), "sigma": (3,), "amplitude": 5})
gain = StaticGain("gain", {"factor": 0.8})
nf = NeuralField("field", {
    "shape": (50,), "resting_level": -5, "global_inhibition": -0.01,
    "tau": 0.1, "input_noise_gain": 0.1,
    "sigmoid": "AbsSigmoid", "beta": 100, "theta": 0.5,
    "LateralKernel": Gaussian({"sigma": (3,), "amplitude": 5, "normalized": True, "max_shape": (50,)}),
})

# Connect steps using >> operator
gi >> gain >> nf

# Compile and run
arch = get_arch()
arch.compile()
for _ in range(100):
    arch.tick()

Documentation Overview

Section	Description
Installation	How to install JUNIPER and its dependencies
Building Architectures	How to create steps, connect them, and run simulations
Command-Line Reference	All run.py arguments
Steps Reference	Complete reference for all built-in steps
Configurables	Configurable objects (Gaussian, LateralKernel, Sigmoid)