Home▸AI & Machine Learning▸Autoencoder Anomaly Detection: Error Heatmap & Latent Trajectory

Autoencoder Anomaly Detection: Error Heatmap & Latent Trajectory

Watch a real tiny autoencoder train live on correlated multi-sensor data. A scrolling per-channel error heatmap and a 2D latent-space trajectory show exactly which sensor drives an anomaly and how the compressed code moves, all on a flat 2D canvas.

AI & Machine Learning2DAdvanced60 FPS📱 Mobile-adapted⇄ 3D version

Details

CategoryAI & Machine Learning

RenderingCanvas 2D

DifficultyAdvanced

Target FPS60

ControlsBottleneck-width slider (2-6) resizes the autoencoder's compression layer and reinitializes its weights; anomaly-injection-rate slider sets how often a corrupted reading enters the stream; corruption-magnitude slider sets how far affected channels are pushed off the normal manifold; Training toggle freezes/resumes live gradient-descent learning; Reset counters clears the detected/injected anomaly tally.

autoencoderanomaly detectionreconstruction errorneural networkunsupervised learningsensor monitoringlatent spaceheatmap

Related simulations

Autoencoder Anomaly Detection: Reconstruction Error SimulatorStream data through an encoder-decoder neural network, watch it get squeezed through a…

Autoencoder Reconstruction Error Anomaly Detection LabA 3D lab showing how an autoencoder learns a 'normal' data manifold and flags anomalies by…

Artificial Intelligence Explained: Neural Network Learning SimulatorWatch a small feedforward neural network learn in real time. Switch between supervised,… ⊞ All 1161 in AI & Machine Learning →

The same real 8-channel autoencoder as the 3D version — a tiny encoder-decoder neural network trained live by gradient descent — but rendered entirely as flat 2D plots instead of a 3D bar scene. A scrolling per-channel error heatmap shows which sensor is driving an anomaly the instant it happens, while a 2D phase portrait of the compressed latent code reveals the hidden two-factor structure the network discovered: normal samples trace a closed loop, anomalies kick the trajectory off it. Tune the bottleneck width to see the classic compression/detectability trade-off, inject anomalies at different rates and magnitudes, and freeze training to compare a converged network against one still learning.

⚙ Under the hood

A real 8-channel autoencoder trains live by gradient descent on correlated sensor data, then reconstructs each new reading; a scrolling per-channel error heatmap and a 2D latent-space phase portrait pinpoint exactly which sensor is driving an anomaly and how the compressed code moves, all on a flat 2D canvas.

autoencoderanomaly detectionreconstruction errorneural networkunsupervised learningsensor monitoringlatent spaceheatmap

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install