DYLATIS

We have recently begun development of DYLATIS — the Dynamic Latent Taxonomy Identification Space, a framework created to answer a central question in the OFELIA project:

What level of recognition accuracy can we expect when more than thirty deep-learning–based feature-detector models — each with known or estimated accuracy — operate together within a vast, noisy, high-dimensional species universe?

Early experimental phases of OFELIA revealed an uncomfortable truth. Methods that behave well at small scale — a few hundred species (<1000) — collapse completely once we cross the threshold of ~10,000 species. Traditional flat classification, simple distance metrics, and textbook feature-matching pipelines remain usable for small species spaces, but anyone aiming to recognise an entire biological kingdom — the full diversity of fungi, plants, or animals — quickly encounters their fundamental scalability limits.

Complexity does not grow linearly; it grows explosively. Managing thousands of species requires a kingdom-independent architecture built not from ad hoc heuristics, but from general, deeply structured components.

DYLATIS provides exactly this foundation. We construct an artificial species universe directly in latent feature space, where each species is described not by images but by observable feature vectors — traits perceptible to human observers or measurable by machine models. Species may be derived directly from nature (digital twins) or synthetically generated to follow real-world distributional laws.

We first generated synthetic species according to the empirical distributions of key morphological characters, then positioned them relative to one another using inter-species distance patterns observed in nature. Feature generation respects real marginal distributions, entropic structure, clustering morphology, and empirical inter-species distance distributions.

A crucial insight emerged: recognition in this latent taxonomy does not require a classical, metric distance function. It is sufficient to estimate pairwise similarity dimension-by-dimension using the Confusion Distribution Matrix (CDM) associated with each feature-detector model. Each dimension therefore induces its own local notion of “distance”, reflecting how real deep-learning models distort species features. The result is a synthetic taxonomic identification space in which each instance effectively experiences a slightly different distance landscape. In our simulations, we generated a 10⁴-scale synthetic species population.

Individual instances are sampled by applying the pre-assigned intra-species variances along each feature dimension, reproducing natural variability. For each model’s estimated accuracy, we generate its corresponding Confusion Distribution Matrix. This enables DYLATIS to simulate OFELIA’s recognition behaviour well before the full real system exists.

A key innovation is the emergence of latent taxonomy. Instead of imposing genera or families from outside, species self-organise into clusters purely according to the geometry of latent space.

The objective is clear: to predict, with high fidelity, the recognition accuracy OFELIA can achieve when dozens of deep-learning models of known reliability operate jointly on tens of thousands of species.

Through large-scale Monte Carlo simulations in this latent world, DYLATIS produces estimates for top-1, top-5, and cluster-level accuracy under varying noise levels, feature quality conditions, and species counts.

Perhaps most importantly, DYLATIS serves as a guide toward global-scale biological identification. Any system — OFELIA or otherwise — that aims to recognise an entire kingdom must transcend classical techniques.

DYLATIS is our experimental universe for discovering how such a system must behave — and what accuracy can be achieved when we push its limits to the edge.