✦ AE · 3→64→64→2→64→64→3 · ELU · MSE

How Autoencoders Work

A real neural network (backprop + AdamW) learns to compress 3D space into a 2D latent space. Every reset draws new random weights — convergence path differs every run. Customize the architecture below.

Recon Loss

—

Grad Steps

KL × β

—

Training Control

Step 0 — Random weightsStep 3000 — Converged

⚙ Model Configuration ▼

Input Space

3D structural descriptors (fixed)

⟳ rotating

Latent Space — Encoder

2D codes from actual encoder forward pass

Reconstructed Output

Ghost=input · dots=reconstruction · color=error magnitude

⟳ rotating

low error → high error

Training Loss Curve

Real MSE loss computed from actual forward pass each step