Latent Space Exploration - Seeds Of Hawaii

Latent Space Exploration Project

Imagine you are the botanical collections manager for a museum. On your first day, the head of research asks you to start cataloging plants that don’t exist. This would surely be a first. Museums deal in real world specimens with provenance—not speculation. Asking museum staff to catalog things in their imagination is like putting together an expedition to find a unicorn.

But assuming you wanted to keep your job, how would you proceed? What methods would you employ to complete such a quixotic task?

Every texture, color combo, branching pattern, and chemical trick that evolution has cooked up—plus millions more it hasn’t gotten around to yet—would somehow need to be systematically documented. That’s the ridiculous multidimensional maze of plant possibilities you’d be stuck trying to organize.

The effort to represent ALL variations in such a situation would be herculean. Let’s say you wanted to start with every shade of color a plant may possess. You could create thousands of color swatches, write detailed descriptions, build elaborate filing systems. Or you could store it as math.

In the real world, these phenotypic relationships correspond with metadata like taxonomic classifications, geographic distributions, and ecological associations. But instead of drowning in endless spreadsheets and filing cabinets, latent space compresses the storage method—not the information itself.

Latent space doesn’t simplify complexity away; it maps every nuance into a mathematical realm where we can navigate and explore it systematically. It’s not about reducing complexity, but about making infinite complexity explorable.

But why would you want to explore this mathematical space? Consider two real seed specimens from the Lyon Arboretum Seed Lab at the University of Hawaii: Abutilon eremitopetalum and Abutilon incanum. These are actual seeds, documented and catalogued. But what about the morphological possibilities that exist between them? What seed forms could theoretically bridge these different expressions of the same genus?

Traditional seed banking stops at the specimens you can physically collect and store. But latent space lets you explore the ‘what-ifs’—the morphological possibilities that could exist in nature’s design space. By interpolating between these two Abutilon seeds in mathematical space, you can visualize gradual transitions: shapes slowly shifting from angular to rounded, surfaces becoming more or less textured, proportions subtly adjusting step by step.

This isn’t about reconstructing evolutionary history—it’s about exploring evolutionary potential. Understanding these possible intermediate forms helps reveal the constraints and possibilities within plant development, the range of viable seed designs, and the morphological space that evolution could still explore. It’s like having a map that shows you not just where evolution has been, but where it might go. Suddenly, that impossible task of cataloging seeds that don’t exist becomes a way to understand the boundaries of what’s possible—and the endless forms that nature hasn’t tried yet.

When this approach is grounded in detailed scientific data for each species—rainfall patterns, soil composition, elevation, temperature ranges, geographic coordinates—latent space becomes a powerful tool for visualizing the relationship between environmental pressures and phenotypic variation. Instead of just seeing two different seed forms, you can explore how morphology responds to ecological gradients.

A semantic approach to this process vastly outperforms raw data queries when it comes to visual representations. Imagine you want to explore seeds adapted to moderate rainfall conditions. With raw data, you might linearly interpolate 50% between a desert species receiving 5 inches annually and a rainforest species receiving 60 inches, landing at 32.5 inches. But that mathematical midpoint tells you nothing about what the seed should actually look like.

A semantic latent space, however, understands the complex relationships between all measured dimensions simultaneously. That 50% interpolation between two seed sizes might correspond with 75% between two different rainfall averages, 30% between soil pH levels, and 85% between temperature ranges—because that’s how these environmental factors actually correlate in nature. The model learns that certain combinations of traits co-occur: seeds from arid regions with high UV exposure tend to have thicker, more textured coats, while those from stable rainforest environments can afford thinner, smoother surfaces.

This multidimensional semantic understanding allows for visualizations that interpolate between all possible phenotypic expressions more seamlessly. Rather than crude mathematical averaging across isolated variables, you get morphological predictions that reflect the true complexity of how plants adapt to their environments—revealing forms that make biological sense within the constraints of real-world ecological relationships.

Making Latent Space Visible Through Direct Mapping

While computers can navigate high-dimensional latent spaces through mathematical operations, humans cannot directly perceive these abstract mathematical relationships. However, when we map data dimensions directly to visual features—as demonstrated in this seed interpolation network—we suddenly gain intuitive access to latent space exploration.

Above is a rudimentary visualization, designed to show each seed’s position in 2D space. Each seeds owes its location in multi-dimensional feature space to morphological characteristics (length, width, mass, texture). The original species occupy the corners of this space, while the interpolated specimens lie along the connecting paths, showing smooth transitions between parent characteristics.

By clicking through the network, you can literally see how features blend and morph—the visual representation makes the abstract mathematical interpolation tangible.

This direct visual-to-data mapping transforms latent space from an incomprehensible mathematical abstraction into an explorable landscape. Manipulating vectors in high-dimensional space is aided through visual intuition, allowing you to understand relationships and discover meaningful variations. The network reveals not just individual specimens, but the underlying structure of morphological possibility space—exactly the kind of visual latent space exploration we’re building for creative and scientific applications.

This is precisely what makes visual latent space interfaces so powerful: they leverage human visual pattern recognition to make complex mathematical relationships immediately comprehensible and navigable.

Computers can navigate high-dimensional latent spaces through mathematical operations, but humans cannot directly perceive these abstract mathematical relationships. However, when data dimensions are mapped directly to visual features, latent space becomes intuitively accessible.

In biological morphology, specimens exist in multi-dimensional feature spaces defined by characteristics like length, width, mass, and texture. Original species occupy distinct points in this space, while interpolated specimens represent smooth transitions between parent characteristics, showing how features blend and morph mathematically.

This direct visual-to-data mapping transforms latent space from an incomprehensible mathematical abstraction into an explorable concept. Instead of manipulating vectors in high-dimensional space, visual intuition allows us to understand relationships and discover meaningful variations. This reveals not just individual specimens, but the underlying structure of morphological possibility space—exactly the kind of visual latent space exploration we’re building for creative and scientific applications.