sara-engine 0.2.2

A bio-inspired Spiking Neural Network engine with Hugging Face-like pipeline APIs (CPU-only, Backprop-free).

SARA Engine (Liquid Harmony)

SARA (Spiking Advanced Recursive Architecture) is a next-generation AI engine that mimics the biological brain's power efficiency and event-driven processing.

It completely eliminates the "backpropagation (BP)" and "matrix operations" that modern deep learning (ANNs) rely on, achieving advanced recognition, text generation, and retrieval capabilities using only sparse spike communication.

It operates completely on CPU, without using any GPU or NumPy.

Current Version: v0.2.2

Features

• Hugging Face-like Pipelines: Incredibly easy-to-use API (pipeline("text-generation"), pipeline("image-classification"), etc.).
• No Backpropagation: Learns natively using biological rules like Spike-Timing-Dependent Plasticity (STDP) and Homeostatic Plasticity.
• Zero Matrix Math: Replaces dense tensor multiplications with purely discrete spike routing.
• CPU Only & Eco-Friendly: Does not require expensive GPU resources.

Installation

pip install sara-engine

Quick Start

SARA v0.2.2 introduces an incredibly intuitive pipeline API, bringing the ease of modern NLP frameworks to Spiking Neural Networks.

1. Text Generation (Zero-Shot Inference)

You can easily load pre-trained STDP synapses and generate text autoregressively.

from sara_engine import pipeline

# Load a biologically trained model and tokenizer
generator = pipeline("text-generation", model="path/to/saved_snn_model")

output = generator("Hello, I am a spiking", max_new_tokens=15)
print(output[0]['generated_text'])

2. Feature Extraction & RAG

Extract semantic embeddings using Liquid State Machines (LSM) without any mathematical attention layers.

from sara_engine import pipeline

extractor = pipeline("feature-extraction", model="path/to/saved_extractor")

# Extracts a high-dimensional membrane potential vector
vector = extractor("Artificial intelligence is evolving.")
print(f"Spike Vector Length: {len(vector)}")

3. Image Classification

Process pixels via Retinal Rate Coding directly into SNN layers.

from sara_engine import pipeline

vision_classifier = pipeline("image-classification", model="path/to/vision_model")

# Pass a 2D array of pixel intensities (0.0 to 1.0)
image = [
    [1.0, 0.0, 1.0],
    [0.0, 1.0, 0.0],
    [1.0, 0.0, 1.0]
]
prediction = vision_classifier(image)
print(prediction[0]['label']) # e.g., "Cross (X)"

License

MIT License