Interactive 3D visualization of how neural network embeddings evolve across layers

##🧩 Conclusion

This 3D visualization reveals how a transformer model energizes and organizes meaning across its layers. Initially, all sentence embeddings cluster tightly — representing low semantic differentiation. As the layers deepen, the embeddings spread outward and form distinct clusters, showing how the model gradually transforms raw token information into semantically charged representations (positive, negative, neutral).

The expanding, conical structure reflects a flow of informational energy — from a compressed input state to a more expressive, task-specific space. In essence, this plot captures the evolution of understanding inside the model, where representational energy becomes more structured and sentiment meaning emerges layer by layer.

@article{yadav{{ page.date | date: "%Y" }}{{ page.title | slugify | replace: '-', '' }},
    title = {layers-travel},
    author = {Sumit Yadav},
    journal = {Tatva},
    year = {{{ page.date | date: "%Y" }}},
    month = {{{ page.date | date: "%B" }}},
    day = {{{ page.date | date: "%d" }}},
    url = {https://tatva.sumityadav.com.np/posts/2025/09/16/layers-travel/},
    note = {Accessed: {{ site.time | date: "%B %d, %Y" }}}
}

Yadav, S. ({{ page.date | date: "%Y, %B %d" }}). layers-travel. ." Tatva, {{ page.date | date: "%d %b %Y" }}, https://tatva.sumityadav.com.np." Tatva. {{ page.date | date: "%B %d, %Y" }}. https://tatva.sumityadav.com.np/posts/2025/09/16/layers-travel/.