// bridge system

useState()→Model Weights·Event Propagation→Forward Pass·Array.map()→Tensor Operation·React diff→Loss Function L = Σ(y−ŷ)²·transition-duration→Learning Rate η·CSS clamp()→σ(x) Activation·Re-render cycle→Training Epoch·Event bubbling→Backpropagation ∂L/∂w·useCallback→Gradient Caching·Promise.all()→Batch Inference·Redux store→Weight Matrix·DevTools profiler→Loss Landscape·useState()→Model Weights·Event Propagation→Forward Pass·Array.map()→Tensor Operation·React diff→Loss Function L = Σ(y−ŷ)²·transition-duration→Learning Rate η·CSS clamp()→σ(x) Activation·Re-render cycle→Training Epoch·Event bubbling→Backpropagation ∂L/∂w·useCallback→Gradient Caching·Promise.all()→Batch Inference·Redux store→Weight Matrix·DevTools profiler→Loss Landscape·

Rohan
Listen to this. Every note in this soundscape was sampled from a probability distribution I designed. High notes are rare — low probability. Bass tones are common — high probability. The distribution is the composition. Sampling is the performance.

MUSE
To sample is to reach into possibility and pull out a single truth.

In frontend development, you call Math.random() and get a number. That is sampling from a uniform distribution. But generative AI needs more control — sampling from specific distributions, at specific temperatures, with reproducible seeds.

Learning Objectives

○Understand sampling as drawing values from distributions
○Implement different sampling strategies in TensorFlow.js
○Control randomness with temperature scaling
○Use seeded randomness for reproducible generation

From Distribution to Data Point

Sampling bridges the gap between an abstract distribution and a concrete value. When a generative model runs, its final act is always sampling.

Frontend

Seeded Random

function seededRandom(seed: number) { /* deterministic */ }

Machine Learning

Sampling Strategy

const sample = tf.randomNormal([1, 100], 0, 1)

Structural Bridge

sampling-basics.tstypescript

import * as tf from '@tensorflow/tfjs';

// Basic sampling: draw from standard normal
const samples = tf.randomNormal([5]);
console.log(await samples.array());
// e.g., [-0.42, 1.13, -0.87, 0.21, 0.65]

// Transformed sampling: shift and scale
const mean = 100;
const stdDev = 15;
const iqSamples = tf.randomNormal([1000]).mul(stdDev).add(mean);
// Now samples cluster around 100 with spread of 15

// Frontend parallel: weighted random selection
function weightedChoice<T>(items: T[], weights: number[]): T {
const total = weights.reduce((a, b) => a + b, 0);
let r = Math.random() * total;
for (let i = 0; i < items.length; i++) {
  r -= weights[i];
  if (r <= 0) return items[i];
}
return items[items.length - 1];
}
// This IS sampling from a categorical distribution

Temperature: Controlling Creativity

Temperature is a scaling factor that controls how "creative" or "conservative" sampling is. Low temperature concentrates probability on the most likely values. High temperature spreads it out, allowing wilder outputs.

temperature-sampling.tstypescript

import * as tf from '@tensorflow/tfjs';

// Logits: raw model output (unnormalized scores)
const logits = tf.tensor1d([2.0, 1.0, 0.5, -1.0]);

function sampleWithTemperature(logits: tf.Tensor1D, temp: number) {
// Divide logits by temperature before softmax
const scaled = logits.div(temp);
const probs = scaled.softmax();
return probs;
}

// Low temperature (0.1): almost deterministic
// → [0.9999, 0.0001, 0.0000, 0.0000]
const conservative = sampleWithTemperature(logits, 0.1);

// Temperature 1.0: standard softmax
// → [0.506, 0.186, 0.113, 0.025]
const standard = sampleWithTemperature(logits, 1.0);

// High temperature (5.0): nearly uniform
// → [0.314, 0.268, 0.249, 0.169]
const creative = sampleWithTemperature(logits, 5.0);

Think of temperature like a thermostat for creativity. Rohan uses it in his audio: low temperature for predictable rhythms, high temperature for experimental noise.

Challenge

⚡

Exercise

Beginner~8 min

Temperature-Controlled Sampling

Implement a function that applies temperature scaling to logits and returns the resulting probability distribution. Divide the logits by the temperature, then apply softmax.

import * as tf from '@tensorflow/tfjs';

async function sampleWithTemperature(
  logits: number[],
  temperature: number
): Promise<number[]> {
  const logitsTensor = tf.tensor1d(logits);
  // 1. Divide logits by temperature
  // 2. Apply softmax to get probabilities
  const probs = null; // your code here
  return probs;
}

Key Takeaways

✓Sampling draws concrete values from abstract probability distributions
✓Temperature controls the randomness-creativity tradeoff
✓Low temperature produces conservative output; high temperature produces diverse output
✓Seeded randomness enables reproducible generation for debugging and testing

Need a hint?

🧭 Guidance

✅ Solution

Sampling — Nova Canvas | Tensorcraft

Sampling

Context