NOTES FROM CLASS

• we know how to go from text to embedding so words can be compared mathematically. What if we apply operations to those embeddings, how do we go back to text?
  • How can we simulate this, even with a shortcut in LLM
  • I dont know of a model that goes forwards and backwards
  • I have a horse and a duck, i tack the average of those concepts in an embedding. But if i take that average how do i go back to a conceptual animal that
  • If you have database of descriptions, you could try to get an LLM to take the average of descriptions
  • You could try to embed the genomes of the animals. That could be a cool visualization.
  • GANs + Prompting / descriptions — maybe this would ruin the project conceptually, but it would certainly work: you can easily get an image generator to do exactly what were talking about. Two prompts: draw a horse, draw a turtle, average the turtle. Then give them a percentage, then interpoliate them, then as a vision model to describe it. Use an image generator to do the embeddings math
  • Because there isn’t anything that exists, you can decide
• Embeddings as mapped in 2D dont feel meaningful. UMAP is just doing its best. The higher dimensional data is where th emeaning is. Having descriptions as opposed to just single words. The semantic meaning will give me more meaning. Make dataset richer in UMAP

An interactive visualization tool that uses language models to invent plausible new animals positioned conceptually between known animals in a semantic embedding space.

Screen Recording 2025-04-18 at 8.19.47 AM.mov

Vision and Concept

Core Idea

Language is a dynamic probability space where familiar words represent known points, and new words emerge naturally to fill semantic gaps. By visualizing this space, we can explore how concepts evolve and new ideas are formed.

Goal

Create an interactive two-dimensional map of animals derived from word embeddings, allowing users to click between known animal points to generate new, plausible animals. This visually demonstrates how language evolves through creative interpolation between existing ideas.

Inspirational Sources

• Quotes
  • “All language is a set of labels placed on invisible concepts.” – Ferdinand de Saussure
  • "The limits of my language mean the limits of my world." – Ludwig Wittgenstein
• Projects
  • Allison Parrish’s "Understanding Word Vectors"
  • Wordnet Database: https://wordnet.princeton.edu
  • Embedding Projector: https://projector.tensorflow.org
• Images
  • Historical bestiaries (medieval illustrations of hybrid creatures)
  • Borges' fictional animal compendium from The Book of Imaginary Beings
• Conceptual Inspiration
  • Mythological animals as historical interpolations (e.g., unicorn, griffin, hippocampus)
  • Evolutionary biology trees as conceptual maps
  • GANs and Image Interpolation models

Technical Approach

Embedding Space

• Use semantic embeddings to position animal concepts in a meaningful two-dimensional space based on properties such as:
  • Habitat (Aquatic ↔ Terrestrial ↔ Aerial)
  • Morphology (Four-legged ↔ Winged ↔ Fin-bearing)
  • Features (Horned, Scaled, Feathered, etc.)