Skip to workspace

Model Lab

Make the hidden mechanics visible: compression, attention, sampling, and retrieval.

A deterministic linear-bottleneck simulation on correlated 2D points. The “training” control interpolates toward a fixed projection; it is not gradient-based training of a neural autoencoder. An original PyTorch training loop is available in Python Playground.

Autoencoder experiment

Normal reconstruction MSE

0.0018

Noise 0.2 · 40 simulated epochs · 1 latent dimension(s)

Simulated normal reconstruction loss0.350.00StartEnd

Anomaly contrast

Injected sample [0.8, −0.8] has reconstruction error 0.6400.

Illustrative threshold 0.15: Flagged.

Identity capacity can reconstruct the anomaly well and erase the detection signal.

Threshold is illustrative, not calibrated. Noise can create false positives.

Sample
Normal 1
Clean
[-0.95, -0.95]
Noisy
[-1, -1.032]
Latent
[-1.016]
Reconstruction
[-0.98, -0.98]
MSE
0.0016
Sample
Normal 5
Clean
[-0.55, -0.55]
Noisy
[-0.475, -0.451]
Latent
[-0.463]
Reconstruction
[-0.447, -0.447]
MSE
0.0004
Sample
Normal 10
Clean
[-0.05, -0.05]
Noisy
[0.035, 0.028]
Latent
[0.031]
Reconstruction
[0.03, 0.03]
MSE
0.0000
Sample
Normal 15
Clean
[0.45, 0.45]
Noisy
[0.443, 0.362]
Latent
[0.403]
Reconstruction
[0.388, 0.388]
MSE
0.0019
Sample
Normal 20
Clean
[0.95, 0.95]
Noisy
[0.992, 0.994]
Latent
[0.993]
Reconstruction
[0.958, 0.958]
MSE
0.0013
Sample
Injected anomaly
Clean
[0.8, -0.8]
Noisy
[0.8, -0.8]
Latent
[0]
Reconstruction
[0, 0]
MSE
0.6400
Read the lesson
Bring this lab to your team Original teaching examples · No production training claims