Experience Neuraxon's trinary neural dynamics with our interactive 3D visualization at HuggingFace Spaces.
- π§ Build Custom Networks: Configure neurons, synapses, and plasticity parameters
- π― Interactive Controls: Manually set input neuron states (excitatory/neutral/inhibitory)
- π¬ Live Neuromodulation: Adjust dopamine π―, serotonin π, acetylcholine π‘, and norepinephrine β‘ in real-time
- π 3D Visualization: Watch neural activity flow through the network with curved synaptic connections
- βοΈ Preset Configurations: Try small networks, large networks, high plasticity modes, and more
βΆοΈ Real-time Simulation: Run continuous processing and observe emergent dynamics
No installation requiredβjust open your browser and explore!
Interactive 3D visualization showing neural activity and neuromodulator flow
Neuraxon is a bio-inspired neural network framework that extends beyond traditional perceptrons through trinary logic (-1, 0, 1), capturing excitatory, neutral, and inhibitory dynamics found in biological neurons.
Unlike conventional neural networks that use discrete time steps and binary activation, Neuraxon features:
- Continuous processing where inputs flow as constant streams
- Multi-timescale computation at both neuron and synapse levels
- Dynamic plasticity with synaptic formation, collapse, and rare neuron death
- Neuromodulation inspired by dopamine, serotonin, acetylcholine, and norepinephrine
- Spontaneous activity mirroring task-irrelevant yet persistent brain processes
This implementation includes a hybridization with Qubic's Aigarth Intelligent Tissue, demonstrating evolutionary approaches to neural computation.
Check out our paper for complete theoretical foundations and biological inspirations!
Neuraxons operate in three states:
- +1 (Excitatory): Active firing, promoting downstream activity
- 0 (Neutral): Subthreshold processing, enabling subtle modulation
- -1 (Inhibitory): Active suppression of downstream activity
This third "neutral" state models:
- Metabotropic receptor activation
- Silent synapses that can be "unsilenced"
- Subthreshold dendritic integration
- Neuromodulatory influences
Each synapse maintains three dynamic weights:
w_fast # Ionotropic (AMPA-like), Ο ~5ms - rapid signaling
w_slow # NMDA-like, Ο ~50ms - sustained integration
w_meta # Metabotropic, Ο ~1000ms - long-term modulationUnlike discrete time-step models, Neuraxon processes information continuously:
Ο (ds/dt) = -s + Ξ£ w_iΒ·f(s_i) + I_ext(t)
This enables:
- Real-time adaptation to streaming inputs
- Natural temporal pattern recognition
- Biologically plausible dynamics
git clone https://github.com/DavidVivancos/Neuraxon.git
cd Neuraxon
pip install -r requirements.txtfrom neuraxon import NeuraxonNetwork, NetworkParameters
# Create network with default biologically-plausible parameters
params = NetworkParameters(
num_input_neurons=5,
num_hidden_neurons=20,
num_output_neurons=5
)
network = NeuraxonNetwork(params)
# Set input pattern (trinary states: -1, 0, 1)
network.set_input_states([1, -1, 0, 1, -1])
# Run continuous simulation
for step in range(100):
network.simulate_step()
if step % 20 == 0:
outputs = network.get_output_states()
print(f"Step {step}: Outputs = {outputs}")
# Modulate network behavior via neuromodulators
network.modulate('dopamine', 0.8) # Enhance learning
network.modulate('serotonin', 0.6) # Adjust plasticity
# Save network state
from neuraxon import save_network
save_network(network, "my_network.json")Input Layer (5 neurons)
β β (bidirectional ring connectivity)
Hidden Layer (20 neurons)
β β (with spontaneous activity)
Output Layer (5 neurons)
Constraints:
- Small-world connectivity (~5% connection probability)
- No output β input connections
- Dynamic topology via structural plasticity
Neuraxon implements continuous weight evolution inspired by STDP:
# Weights evolve based on pre/post activity and neuromodulators
# LTP: pre=1, post=1 β strengthen synapse
# LTD: pre=1, post=-1 β weaken synapse
# Neutral state provides nuanced control# Synapses can form, strengthen, weaken, or die
# Neurons can die if health drops below threshold (hidden layer only)
# Silent synapses can be "unsilenced" through correlated activity# Four neuromodulators with distinct roles:
neuromodulators = {
'dopamine': 0.1, # Learning & reward
'serotonin': 0.1, # Mood & plasticity
'acetylcholine': 0.1, # Attention & arousal
'norepinephrine': 0.1 # Alertness & stress response
}Neuraxon is particularly suited for:
- Continuous learning systems that adapt in real-time
- Temporal pattern recognition in streaming data
- Embodied AI and robotics requiring bio-realistic control
- Adaptive signal processing with non-stationary inputs
- Cognitive modeling of brain-like computation
- Energy-efficient AI leveraging sparse, event-driven processing
Visit our HuggingFace Space for a fully interactive 3D visualization where you can:
- Configure all network parameters through an intuitive GUI
- Visualize neurons color-coded by type (Blue input, pink mid, red output) and state:
- High Intesity = Excitatory (+1)
- Mid Intesity = Inhibitory (-1)
- Dark = Neutral (0)
- Watch neuromodulator particles (emoji sprites) flow along synaptic pathways
- Control input patterns and observe how they propagate through the network
- Experiment with different neuromodulator levels and see their effects
- Compare preset configurations (minimal, balanced, highly plastic, etc.)
The demo features a 3D sphere layout with curved synaptic connections and real-time particle effects representing neuromodulator dynamics.
All parameters have biologically plausible default ranges:
@dataclass
class NetworkParameters:
# Architecture
num_input_neurons: int = 5 # [1, 100]
num_hidden_neurons: int = 20 # [1, 1000]
num_output_neurons: int = 5 # [1, 100]
connection_probability: float = 0.05 # [0.0, 1.0]
# Neuron dynamics
membrane_time_constant: float = 20.0 # ms [5.0, 50.0]
firing_threshold_excitatory: float = 1.0 # [0.5, 2.0]
firing_threshold_inhibitory: float = -1.0 # [-2.0, -0.5]
# Synaptic timescales
tau_fast: float = 5.0 # ms [1.0, 10.0]
tau_slow: float = 50.0 # ms [20.0, 100.0]
tau_meta: float = 1000.0 # ms [500.0, 5000.0]
# Plasticity
learning_rate: float = 0.01 # [0.0, 0.1]
stdp_window: float = 20.0 # ms [10.0, 50.0]
# ... see code for complete parameter setThis implementation hybridizes Neuraxon with Aigarth Intelligent Tissue, combining:
- Neuraxon: Sophisticated synaptic dynamics and continuous processing
- Aigarth: Evolutionary framework with mutation and natural selection
The hybrid creates "living neural tissue" that:
- Evolves structure through genetic-like mutations
- Adapts weights through synaptic plasticity
- Undergoes selection based on task performance
- Exhibits emergent complexity and self-organization
If you use Neuraxon in your research, please cite:
@article{Vivancos-Sanchez-2025neuraxon,
title={Neuraxon: A New Neural Growth \& Computation Blueprint},
author={David Vivancos and Jose Sanchez},
year={2025},
journal={ResearchGate Preprint},
institution={Artificiology Research, UNIR University, Qubic Science},
url={https://www.researchgate.net/publication/397331336_Neuraxon}
}We welcome contributions! Areas of interest include:
- Novel plasticity mechanisms
- Additional neuromodulator systems
- Energy efficiency optimizations
- New application domains
- Visualization tools
- Performance benchmarks
Please open an issue to discuss major changes before submitting PRs.
David Vivancos
Artificiology Research https://artificiology.com/ , Qubic https://qubic.org/ Science Advisor
Email: vivancos@vivancos.com
Jose Sanchez
UNIR University , Qubic https://qubic.org/ Science Advisor
Email: jose.sanchezgarcia@unir.net
MIT License. See LICENSE file for details.
Core Neuraxon: Licensed under MIT License (permissive, no restrictions)
Aigarth Hybrid Features: If you implement the Aigarth hybrid features described in our paper, you MUST comply with the Aigarth License, which includes:
- β NO military use of any kind
- β NO use by military-affiliated entities
- β NO dual-use applications with military potential
See NOTICE for full details.
The standalone Neuraxon implementation (without Aigarth integration) has no such restrictions.
This work builds upon decades of neuroscience research on:
- Synaptic plasticity (Bi & Poo, 1998)
- Neuromodulation (Brzosko et al., 2019)
- Spontaneous neural activity (Northoff, 2018)
- Continuous-time neural computation (Gerstner et al., 2014)
Special thanks to the Qubic's Aigarth team for the evolutionary tissue framework integration.