Fuzzy ART

A high-performance implementation of the Fuzzy Adaptive Resonance Theory (ART) algorithm with hardware acceleration support.

Features

• Unsupervised Learning: Efficiently learns patterns with a single data pass
• Online Learning: Enables simultaneous learning and inference without retraining
• Stability/Plasticity: Preserves previously learned information (no catastrophic forgetting)

Performance

• Hardware-Accelerated Computation: Automatically detects and uses the fastest available SIMD instructions
  • Apple Silicon: Leverages the Accelerate framework on macOS/ARM64 platforms (M1/M2/M3 chips)
  • x86 Processors: Utilizes AVX-512 instructions on compatible CPUs
  • Fallback Support: Provides optimized generic implementation for all other systems
• Parallel Processing: Multi-threaded implementation scales with available CPU cores (using a fixed-size worker-pool)
  • Category activation calculations
  • Fuzzy intersection computation
• Memory Efficiency: Reuses pre-allocated resources to minimize garbage collection overhead

Training on the full MNIST dataset completes in 5-6 minutes on Apple M1 Pro / Xeon W-3265M.

Installation

go get -u github.com/oblq/art

Basic Usage

package main

import (
	"fmt"

	"github.com/oblq/art"
)

func main() {
	// Create a new Fuzzy ART model with:
	// - 5 input features
	// - 0.9 vigilance parameter (controls category granularity)
	// - 0.01 choice parameter (influences category competition)
	// - 1.0 learning rate (controls weight update speed)
	model := art.NewFuzzyART(5, 0.9, 0.01, 1)
	defer model.Close() // Release resources when done

	// Prepare an input sample (values should be normalized between 0-1)
	input := []float64{0.1, 0.2, 0.3, 0.4, 0.5}

	// Train the model with the sample
	_, categoryIndex := model.Fit(input)

	// Test the model (with learning disabled)
	resonance, predictedCategoryIndex := model.Predict(input, false)

	fmt.Printf("Matched: %t (Category: %d, Resonance: %.4f)\n",
		categoryIndex == predictedCategoryIndex,
		predictedCategoryIndex,
		resonance)
}

MNIST Example

The repository includes a full example using the MNIST dataset.

Downloading the Dataset

make get-mnist

This command downloads the MNIST training and test sets to the testdata directory.

Running the Example

make run

This trains and tests the Fuzzy ART model on the MNIST dataset, automatically using the optimal hardware acceleration for your system.

About Adaptive Resonance Theory

Adaptive Resonance Theory (ART) is a cognitive and neural theory developed by Stephen Grossberg and Gail Carpenter that explains how the brain autonomously learns to categorize, recognize, and predict objects and events in a changing environment.

For a comprehensive overview, refer to this research paper.