Latent-Space Results
Load example data or enter your own dataset.
Valid samples
0
Latent dimensions
0
Classes
0
Clusters
0
Silhouette
—
Anomalies
0
Mean error
—
Highest error
—
Davies–Bouldin
—
Calinski–Harabasz
—
Explained variance
—
Projection
—
Use box or lasso selection to inspect samples. Double-click the plot to reset axes.
Reconstruction Error Distribution
Processed Samples
| ID | Class | Split | X | Y | Z | Cluster | Error | Anomaly |
|---|---|---|---|---|---|---|---|---|
| No processed data. | ||||||||
Interpretation
No analysis has been generated yet.
Formula Used
Reconstruction error
MSE = (1/n) Σ(xᵢ − x̂ᵢ)²
Variational autoencoder objective
L = Lreconstruction + β DKL(q(z|x) || p(z))
Linear interpolation
z(t) = (1 − t)za + tzb
Silhouette coefficient
s(i) = (b(i) − a(i)) / max(a(i), b(i))
How to Use
- Paste latent vectors or upload a supported data file.
- Parse the dataset and review detected columns.
- Select latent dimensions, labels, identifiers, and errors.
- Choose projection, clustering, and anomaly settings.
- Customize the plot and generate the results.
- Inspect points, metrics, clusters, and reconstruction errors.
- Export the chart, data, settings, or PDF report.
Example Data
| sample_id | label | split | z1 | z2 | z3 | reconstruction_error |
|---|---|---|---|---|---|---|
| S001 | Class A | train | -1.82 | -1.14 | 0.21 | 0.031 |
| S006 | Class B | train | 1.68 | 1.02 | -0.31 | 0.027 |
| S011 | Class C | test | -0.11 | 1.61 | 0.62 | 0.058 |
| S013 | Anomaly | test | 3.95 | -3.48 | 2.20 | 0.380 |
Interpretation Guide
- Tight groups indicate similar encoded representations.
- Separated classes suggest useful latent discrimination.
- Overlapping groups may indicate weak representation separation.
- High reconstruction errors may reveal anomalies or rare samples.
- PCA preserves variance but may hide nonlinear structure.
- Projection views never replace analysis in full dimensions.
Frequently Asked Questions
What is an autoencoder latent space?
It is the compressed representation learned by the encoder.
Why use PCA before plotting?
PCA converts many latent dimensions into interpretable visual axes.
What does reconstruction error show?
It measures differences between original and reconstructed samples.
Can this calculator detect anomalies?
Yes. It supports error, score, distance, and percentile thresholds.
What is a good silhouette score?
Values nearer one usually indicate clearer cluster separation.
Does a two-dimensional plot preserve everything?
No. Every projection can lose important high-dimensional relationships.
Can I compare training and testing samples?
Yes. Map a split column and select split coloring.
Does it support variational autoencoders?
Yes. VAE fields and beta-weighted analysis options are included.
Are uploaded files sent to a server?
No. Browser processing keeps uploaded data on your device.