44 lines
1.2 KiB
Markdown
44 lines
1.2 KiB
Markdown
# NeuroGraph
|
|
|
|
Bio-inspired energy-gradient deep learning framework.
|
|
|
|
## Overview
|
|
|
|
NeuroGraph explores training paradigms beyond backpropagation:
|
|
- **Energy-based learning**: Networks relax to energy minima instead of computing analytic gradients
|
|
- **Reward-modulated plasticity**: Three-factor learning rules with neuromodulator signals
|
|
- **Autonomous pruning**: Self-organizing network topology through structural plasticity
|
|
- **Architecture search**: Differentiable graph exploration for optimal connectivity
|
|
|
|
## Setup
|
|
|
|
```bash
|
|
pip install -e ".[dev]"
|
|
```
|
|
|
|
For GPU support:
|
|
```bash
|
|
pip install -e ".[dev,gpu]"
|
|
```
|
|
|
|
## Quick Start
|
|
|
|
```python
|
|
from neurograph.core.energy import compute_energy, EnergyConfig
|
|
|
|
config = EnergyConfig(data_weight=1.0, reg_weight=0.01)
|
|
energy = compute_energy(params, activities, inputs, targets, config=config)
|
|
```
|
|
|
|
## Project Structure
|
|
|
|
```
|
|
src/neurograph/
|
|
├── core/ # Energy functions, equilibrium dynamics, neurons
|
|
├── learning/ # EqProp, reward modulation, Hebbian rules
|
|
├── pruning/ # Magnitude/activity-based pruning, structural plasticity
|
|
├── architecture/ # Graph NAS, topology mutation
|
|
├── env/ # Gymnasium wrappers
|
|
└── utils/ # Visualization, logging, metrics
|
|
```
|