Visualizing Word Embeddings as Discrete and Continuous Structures in VR

​

Understanding high-dimensional data using passthrough AR

The Challenge: Understanding High-Dimensional Data

Machine learning models generate embeddings in hundreds or thousands of dimensions. Traditional visualization flattens these into 2D plots using PCA, t-SNE, or UMAP, which loses critical spatial relationships and structural information that exists in the original high-dimensional space.

​

I study how people understand word embeddings in VR.

​

In my first project, I visualized 64 word embeddings as 2D and 3D point clouds across four categories like emotions and professions.

​

I found that 2D plots gave clearer clusters, while VR improved spatial intuition but still treated meaning as discrete groups.

Transition + Project 2 system

1

But this revealed a limitation — point clouds show where clusters are, but not how meaning changes between them.

2

So in my second project, I reframe embeddings as a continuous semantic field using density and gradients.

3

I built a VR system where flow lines show semantic direction and users can probe local category composition.

Results and Insights

I found that VR fields and flows reveal smoother transitions and more overlap between categories, especially in abstract concepts.

​

Compared to point clouds, this exposes semantic continuity that clustering hides.

​

Overall, the projects show that moving from discrete to continuous representations changes not just visualization, but how people interpret meaning in embedding space.

Come Try Them Out!