Neon Protocol

This sample needs array indexing and weighted sums. Continue directly if those are familiar. Review only the part you need:

• Array shape and axis meaning
• Weighted sums

Spatial neighborhoods and shared kernel weights are the new idea. Finishing Deep Orbit does not establish your ability to evaluate a vision model.

// One output at row r, column c:
let value = 0;
for (let kr = 0; kr < 3; kr++) {
  for (let kc = 0; kc < 3; kc++) {
    value += image[r + kr][c + kc] * kernel[kr][kc];
  }
}

// Equivalent feature map with TensorFlow.js imported as tf:
const values = tf.tidy(() => {
  const input = tf.tensor4d(image.flat(), [1, 5, 5, 1]);
  const filter = tf.tensor4d(kernel.flat(), [3, 3, 1, 1]);
  return Array.from(tf.conv2d(input, filter, 1, 'valid').dataSync());
});

Stride 1 moves one pixel at a time. With no padding, a 3 × 3 kernel fits in three positions along each axis of a 5 × 5 image. CNN libraries call this convolution, although the unflipped-kernel operation is cross-correlation. Here the weights are fixed; a trained convolutional layer learns its weights.

A feature response is not an object label or a probability. The full vision curriculum adds training, evaluation, boxes and masks. This isolated sample does not train a model or record course completion.