Feature Explorer

Adjust VAE dimensions and watch decoded tablet silhouettes change in real time

🎛️ Interactive tool — runs in your browser via Binder

Click Launch Binder below to open a cloud Jupyter session with the trained VAE model pre-loaded. The session takes approximately 2 minutes to start. All computation happens in the cloud — no installation required.

What this tool does

The Feature Explorer (notebook 12. Traverse the VAE latent space.ipynb) lets you:

Select a period and genre from a dropdown (e.g., Ur III — Administrative)
The widget loads a random tablet from that period/genre and displays its mean latent vector as 12 slider values
Adjust any slider to change one VAE dimension while holding all others fixed
Watch the decoded tablet silhouette update in real time
Use this to empirically discover what each dimension physically controls

This is the tool used in the paper to confirm that: - X7 controls height/width ratio (portrait ↔︎ landscape) - X2 controls bounding-box fill efficiency (articulated ↔︎ compact) - X8 controls body unity (segmented ↔︎ slab-like)

Launch

Launch Binder

Or launch the full repository environment and navigate to notebook 12:

Launch full environment

Screenshot

Figure 1: Feature Explorer widget: dropdown selects period/genre, sliders control each VAE dimension, decoded silhouette updates in real time.

(Screenshot placeholder — replace with actual screenshot after first run)

Running locally

If you prefer to run locally with faster response times:

git clone https://github.com/DigitalPasts/ShapingHistory.git
cd ShapingHistory
pip install -r requirements.txt
jupyter notebook "12. Traverse the VAE latent space.ipynb"

The VAE model (~87 MB) will be downloaded automatically from Zenodo on first run.

Suggested experiments

Experiment	What to do	What to look for
Confirm X7 = orientation	Set period to Neo-Assyrian; slide X7 from +4 to −4	Portrait → landscape transition
Confirm X2 = fill efficiency	Set period to Old Babylonian; slide X2 from −5 to +4	Outline regularization + widening
Genre comparison	Same period, switch Administrative ↔︎ Literary	Do the starting shapes differ?
Interpolate Ur III → Achaemenid	Set X2 and X7 to Achaemenid admin means	How many steps to recognizable landscape form?
Explore Middle Elamite	Select Middle Elamite period	Confirm already-landscape in X2

--- title: "Feature Explorer" subtitle: "Adjust VAE dimensions and watch decoded tablet silhouettes change in real time" sidebar: analyses toc: false --- ::: {.interactive-callout} ::: {.callout-title} 🎛️ Interactive tool — runs in your browser via Binder ::: Click **Launch Binder** below to open a cloud Jupyter session with the trained VAE model pre-loaded. The session takes approximately 2 minutes to start. All computation happens in the cloud — no installation required. ::: ## What this tool does The **Feature Explorer** (notebook `12. Traverse the VAE latent space.ipynb`) lets you: 1. **Select a period and genre** from a dropdown (e.g., Ur III — Administrative) 2. The widget loads a random tablet from that period/genre and displays its mean latent vector as 12 slider values 3. **Adjust any slider** to change one VAE dimension while holding all others fixed 4. **Watch the decoded tablet silhouette update** in real time 5. Use this to empirically discover what each dimension physically controls This is the tool used in the paper to confirm that: - **X7** controls height/width ratio (portrait ↔ landscape) - **X2** controls bounding-box fill efficiency (articulated ↔ compact) - **X8** controls body unity (segmented ↔ slab-like) ## Launch [![Launch Binder](https://mybinder.org/badge_logo.svg){.binder-badge}](https://mybinder.org/v2/gh/DigitalPasts/ShapingHistory/HEAD?filepath=12.%20Traverse%20the%20VAE%20latent%20space.ipynb) Or launch the full repository environment and navigate to notebook 12: [![Launch full environment](https://mybinder.org/badge_logo.svg){.binder-badge}](https://mybinder.org/v2/gh/DigitalPasts/ShapingHistory/HEAD) ## Screenshot ![Feature Explorer widget: dropdown selects period/genre, sliders control each VAE dimension, decoded silhouette updates in real time.](../assets/img/widget-explorer-screenshot.png){#fig-explorer-screenshot} *(Screenshot placeholder — replace with actual screenshot after first run)* ## Running locally If you prefer to run locally with faster response times: ```bash git clone https://github.com/DigitalPasts/ShapingHistory.git cd ShapingHistory pip install -r requirements.txt jupyter notebook "12. Traverse the VAE latent space.ipynb" ``` The VAE model (~87 MB) will be downloaded automatically from Zenodo on first run. ## Suggested experiments | Experiment | What to do | What to look for | |---|---|---| | Confirm X7 = orientation | Set period to Neo-Assyrian; slide X7 from +4 to −4 | Portrait → landscape transition | | Confirm X2 = fill efficiency | Set period to Old Babylonian; slide X2 from −5 to +4 | Outline regularization + widening | | Genre comparison | Same period, switch Administrative ↔ Literary | Do the starting shapes differ? | | Interpolate Ur III → Achaemenid | Set X2 and X7 to Achaemenid admin means | How many steps to recognizable landscape form? | | Explore Middle Elamite | Select Middle Elamite period | Confirm already-landscape in X2 |