Merge commit 'd803bfe2b1fe7f5e219e50ac20d6801a0a58ac75' as 'vendor/ruvector'
This commit is contained in:
ruv
171
vendor/ruvector/npm/packages/spiking-neural/examples/pattern-recognition.js
vendored
Normal file
171
vendor/ruvector/npm/packages/spiking-neural/examples/pattern-recognition.js
vendored
Normal file
@@ -0,0 +1,171 @@
|
||||
#!/usr/bin/env node
|
||||
|
||||
/**
|
||||
* Pattern Recognition with Spiking Neural Networks
|
||||
*
|
||||
* This example demonstrates:
|
||||
* - Rate-coded input encoding
|
||||
* - STDP learning (unsupervised)
|
||||
* - Pattern classification
|
||||
* - Lateral inhibition for winner-take-all
|
||||
*/
|
||||
|
||||
const {
|
||||
createFeedforwardSNN,
|
||||
rateEncoding,
|
||||
native,
|
||||
version
|
||||
} = require('spiking-neural');
|
||||
|
||||
console.log(`\nPattern Recognition with SNNs v${version}`);
|
||||
console.log(`Native SIMD: ${native ? 'Enabled' : 'JavaScript fallback'}\n`);
|
||||
console.log('='.repeat(60));
|
||||
|
||||
// Define 5x5 patterns
|
||||
const patterns = {
|
||||
'Cross': [
|
||||
0, 0, 1, 0, 0,
|
||||
0, 0, 1, 0, 0,
|
||||
1, 1, 1, 1, 1,
|
||||
0, 0, 1, 0, 0,
|
||||
0, 0, 1, 0, 0
|
||||
],
|
||||
'Square': [
|
||||
1, 1, 1, 1, 1,
|
||||
1, 0, 0, 0, 1,
|
||||
1, 0, 0, 0, 1,
|
||||
1, 0, 0, 0, 1,
|
||||
1, 1, 1, 1, 1
|
||||
],
|
||||
'Diagonal': [
|
||||
1, 0, 0, 0, 0,
|
||||
0, 1, 0, 0, 0,
|
||||
0, 0, 1, 0, 0,
|
||||
0, 0, 0, 1, 0,
|
||||
0, 0, 0, 0, 1
|
||||
],
|
||||
'X-Shape': [
|
||||
1, 0, 0, 0, 1,
|
||||
0, 1, 0, 1, 0,
|
||||
0, 0, 1, 0, 0,
|
||||
0, 1, 0, 1, 0,
|
||||
1, 0, 0, 0, 1
|
||||
]
|
||||
};
|
||||
|
||||
// Visualize patterns
|
||||
console.log('\nPatterns:\n');
|
||||
for (const [name, pattern] of Object.entries(patterns)) {
|
||||
console.log(`${name}:`);
|
||||
for (let i = 0; i < 5; i++) {
|
||||
const row = pattern.slice(i * 5, (i + 1) * 5).map(v => v ? '##' : ' ').join('');
|
||||
console.log(` ${row}`);
|
||||
}
|
||||
console.log();
|
||||
}
|
||||
|
||||
// Create SNN
|
||||
const n_input = 25; // 5x5 pixels
|
||||
const n_hidden = 20; // Hidden layer
|
||||
const n_output = 4; // 4 pattern classes
|
||||
|
||||
const snn = createFeedforwardSNN([n_input, n_hidden, n_output], {
|
||||
dt: 1.0,
|
||||
tau: 20.0,
|
||||
v_thresh: -50.0,
|
||||
v_reset: -70.0,
|
||||
a_plus: 0.005,
|
||||
a_minus: 0.005,
|
||||
init_weight: 0.3,
|
||||
init_std: 0.1,
|
||||
lateral_inhibition: true,
|
||||
inhibition_strength: 15.0
|
||||
});
|
||||
|
||||
console.log(`Network: ${n_input}-${n_hidden}-${n_output} (${n_input * n_hidden + n_hidden * n_output} synapses)`);
|
||||
console.log(`Learning: STDP (unsupervised)`);
|
||||
|
||||
// Training
|
||||
console.log('\n--- TRAINING ---\n');
|
||||
|
||||
const n_epochs = 5;
|
||||
const presentation_time = 100;
|
||||
const pattern_names = Object.keys(patterns);
|
||||
const pattern_arrays = Object.values(patterns);
|
||||
|
||||
for (let epoch = 0; epoch < n_epochs; epoch++) {
|
||||
let total_spikes = 0;
|
||||
|
||||
for (let p = 0; p < pattern_names.length; p++) {
|
||||
const pattern = pattern_arrays[p];
|
||||
snn.reset();
|
||||
|
||||
for (let t = 0; t < presentation_time; t++) {
|
||||
const input_spikes = rateEncoding(pattern, snn.dt, 100);
|
||||
total_spikes += snn.step(input_spikes);
|
||||
}
|
||||
}
|
||||
|
||||
const stats = snn.getStats();
|
||||
const w = stats.layers[0].synapses;
|
||||
console.log(`Epoch ${epoch + 1}/${n_epochs}: ${total_spikes} spikes, weights: mean=${w.mean.toFixed(3)}`);
|
||||
}
|
||||
|
||||
// Testing
|
||||
console.log('\n--- TESTING ---\n');
|
||||
|
||||
const results = [];
|
||||
for (let p = 0; p < pattern_names.length; p++) {
|
||||
const pattern = pattern_arrays[p];
|
||||
snn.reset();
|
||||
|
||||
const output_activity = new Float32Array(n_output);
|
||||
|
||||
for (let t = 0; t < presentation_time; t++) {
|
||||
const input_spikes = rateEncoding(pattern, snn.dt, 100);
|
||||
snn.step(input_spikes);
|
||||
|
||||
const output = snn.getOutput();
|
||||
for (let i = 0; i < n_output; i++) {
|
||||
output_activity[i] += output[i];
|
||||
}
|
||||
}
|
||||
|
||||
const winner = Array.from(output_activity).indexOf(Math.max(...output_activity));
|
||||
const total = output_activity.reduce((a, b) => a + b, 0);
|
||||
const confidence = total > 0 ? (output_activity[winner] / total * 100) : 0;
|
||||
|
||||
results.push({ pattern: pattern_names[p], winner, confidence });
|
||||
console.log(`${pattern_names[p].padEnd(10)} -> Neuron ${winner} (${confidence.toFixed(1)}% confidence)`);
|
||||
}
|
||||
|
||||
// Noise test
|
||||
console.log('\n--- ROBUSTNESS (20% noise) ---\n');
|
||||
|
||||
function addNoise(pattern, noise_level = 0.2) {
|
||||
return pattern.map(v => Math.random() < noise_level ? 1 - v : v);
|
||||
}
|
||||
|
||||
for (let p = 0; p < pattern_names.length; p++) {
|
||||
const noisy_pattern = addNoise(pattern_arrays[p], 0.2);
|
||||
snn.reset();
|
||||
|
||||
const output_activity = new Float32Array(n_output);
|
||||
|
||||
for (let t = 0; t < presentation_time; t++) {
|
||||
const input_spikes = rateEncoding(noisy_pattern, snn.dt, 100);
|
||||
snn.step(input_spikes);
|
||||
|
||||
const output = snn.getOutput();
|
||||
for (let i = 0; i < n_output; i++) {
|
||||
output_activity[i] += output[i];
|
||||
}
|
||||
}
|
||||
|
||||
const winner = Array.from(output_activity).indexOf(Math.max(...output_activity));
|
||||
const correct = winner === results[p].winner;
|
||||
|
||||
console.log(`${pattern_names[p].padEnd(10)} -> Neuron ${winner} ${correct ? '✓' : '✗'}`);
|
||||
}
|
||||
|
||||
console.log('\nDone!\n');
|
||||
Reference in New Issue
Block a user