I Built a Neural Network You Can Watch Train — Forward Pa...

Every neural-network tutorial I tried threw equations at me before I ever saw what was actually happening. I wanted the reverse: watch the activations flow forward, watch the loss bars shrink, watch backprop push gradients right-to-left across the layers.

So I built it. Here's a neural network that trains itself in front of you 👇

What you're actually seeing

Forward pass — particles flow left → right as activations propagate through the layers.
Loss — bars drop each epoch, and the output neurons glow red → gold → green as the error shrinks.
Backpropagation — particles flow right → left, the gradient returning toward the input.
9 epochs, deterministically timed so every run is identical.

No training data is harmed in the making of this animation — it's a faithful visual model of the phases, built for intuition, not for crunching MNIST.

The stack

React for the phase state machine: idle → forward → loss → backward → done
Framer Motion for the particle and node animations
Web Audio API for synthesized sound (no audio files), tied to the dot movement
A small drift-corrected timing loop so the whole run lands on a fixed wall-clock budget

Three things that were trickier than expected

1. Animate particles with CSS transforms, not SVG attributes

My first version animated each particle's cx/cy. It worked but stuttered. Switching to Framer Motion's x/y (which compile to GPU-friendly CSS transforms) made it buttery:

<motion.circle
  r={4}
  cx={0} cy={0}
  initial={{ x: x1, y: y1, opacity: 0, scale: 0 }}
  animate={{ x: [x1, x2], y: [y1, y2], opacity: [0, 1, 1, 0] }}
  transition={{ duration: 0.65, ease: "easeInOut" }}
/>

2. Backprop has to flow backward

Sounds obvious, but my first pass spawned the backprop particles in the same direction as the forward pass. The fix was just swapping the source/target layer so the dots travel from the deeper layer back toward the input:

// Forward: layer l-1 → l   (left → right)
spawnParticles(l - 1, l, FORWARD_COLOR);

// Backprop: layer l → l-1  (right → left)
spawnParticles(l, l - 1, BACKWARD_COLOR);

Tiny change, huge difference in how "correct" it reads.

3. Make the network show it's learning

Loss bars are fine, but I wanted the network itself to react. So the output nodes are colored by the current loss — the same thresholds as the bars, so the legend stays consistent:

function lossColor(loss) {
  if (loss < 0.15) return GREEN;   // basically trained
  if (loss < 0.40) return GOLD;
  return RED;                       // high error
}

Early epochs glow red; by the end they settle into green. You see the network heal.

Bonus: deterministic timing (so I could record it)

setInterval drift made every recording a slightly different length. I anchored a start timestamp and held each epoch to a fixed budget, correcting drift as it goes:

function waitUntil(targetMs) {
  const remaining = targetMs - (Date.now() - runStart);
  return sleep(Math.max(remaining, 0));
}
// ...end of each epoch:
await waitUntil((epoch + 1) * EPOCH_BUDGET_MS);

Now every run lands on the same total time regardless of frame jitter.

What I learned

Animation isn't decoration — putting the direction of data flow on screen taught me backprop better than any equation did.
Browsers block the Web Audio API until a user gesture, so "start" had to be a real click.
Deterministic timing is underrated: it made the thing recordable and the code simpler.

I'm animating a whole set of these — sorting, Dijkstra, hash tables, binary trees. What algorithm should I visualize next? Drop it in the comments 👇

推荐订阅源

DEV Community