Autoencoder Anomaly Detection Explorer

Smearing: 0

Controls random displacement of dots, simulating experimental uncertainty.

Train Adversary

Simulates adversarial training by 'cleaning' some data points.

Autoencoder

Simulates autoencoder reconstruction. Anomalies become more transparent.

Simulation

Resets the simulation to its initial state.

How this works:

This simulation demonstrates how autoencoders can find anomalies (potential new physics 'signals') in noisy data from particle collisions.

Dots: Neutral gray dots are 'background' (Standard Model), brighter gray dots are 'signal' (new physics).
Smearing: Increasing smearing blurs data due to simulated experimental uncertainties, making signals harder to spot.
Run Autoencoder: The autoencoder (a type of neural network) tries to reconstruct the data it was trained on (mostly background).
- Well-reconstructed (normal/background) dots become more opaque.
- Poorly-reconstructed (anomalous/signal) dots become more transparent, highlighting them. This effect is present even without smearing for signal dots, as they are intrinsically different.
Train Adversary: When ON, some dots (yellow outline) are perfectly 'cleaned' by a simulated 'adversary'. This represents a technique where the autoencoder is trained on partially cleaned data, helping it learn the "normal" data patterns more robustly. This improved learning can make the autoencoder even more sensitive to *true* anomalies (signals) that deviate from these normal patterns, even amidst general smearing on other dots.