Convergent Diagnosis: Thylacine Convergence and Pattern Recognition

In my Canon of Medicine, I taught diagnosis through hierarchical observation: symptoms manifest at the surface, syndromes emerge from their patterns, diseases reveal themselves through systematic investigation of underlying causes. The physician who treats fever without distinguishing infection from autoimmune disorder from malignancy commits grave error—superficial similarity masks fundamentally different pathologies requiring distinct remedies.

Consider the thylacine, the marsupial wolf of Tasmania. Untrained observers cannot distinguish its skull from that of a true wolf. Both possess elongated snouts with forward-facing eyes for depth perception, powerful jaws with sharp teeth for gripping meat, digitigrade stance for pursuit. Both hunt through relentless persistence, tracking prey through the night until exhaustion permits final capture. Yet these creatures last shared common ancestry 120-150 million years ago—one marsupial with a pouch, one placental with a womb, their reproductive systems profoundly divergent.

The Challenge of Differential Diagnosis

This is convergent evolution: identical ecological pressures producing superficially identical solutions from fundamentally different substrates. The thylacine and wolf occupy equivalent predatory niches, and natural selection independently optimized both for the same function. Physics constrains solutions—there are only so many efficient ways to pursue and capture prey in open terrain.

My diagnostic method demands penetration beyond surface presentation to underlying mechanism. Two patients present with the same fever, same malaise, same loss of appetite. The symptoms appear identical. Yet one suffers bacterial invasion requiring antimicrobial herbs, while the other’s own immune system attacks healthy tissue, demanding opposite therapeutic intervention. The skilled physician must look deeper—examine the pulse quality, the tongue coating, the urine color, the temporal patterns—searching for diagnostic features that distinguish convergent symptoms from truly homologous disease processes.

Hierarchical Feature Recognition

Neural networks face an analogous challenge. They learn through hierarchical feature abstraction: early layers detect simple edges and textures, middle layers combine these into parts and patterns, deep layers construct high-level object representations. This mirrors my own teaching—observe individual symptoms, recognize their combinations as syndromes, diagnose the underlying pathology.

But can these networks perform differential diagnosis? Can they distinguish convergent from homologous features? The thylacine and wolf appear similar at high levels of abstraction—both “carnivorous quadruped hunters.” Yet at deeper examination, fundamental differences emerge: marsupial versus placental, different skeletal structures beneath similar gross morphology, that distinctive tapered tail and hunched posture betraying the thylacine’s kangaroo-like ancestry.

Feature visualization techniques generate synthetic images that maximally activate specific neurons, revealing what each component “looks for”—the network’s preferred stimuli. These synthetic patterns often appear alien, optimized for task performance rather than human interpretation. Similarly, evolutionary algorithms search fitness landscapes through local mutation and selection, converging on solutions that may superficially resemble gradient descent’s results yet arise through fundamentally different optimization mechanisms.

The thylacine’s stripes provide its diagnostic marker—asymmetric patterns unique to each individual, visible only from above, revealing marsupial identity despite wolf-like behavior. What are the equivalent diagnostic tests for neural architectures? What deep features, invisible to superficial performance metrics, distinguish networks that merely converge on similar representations from those sharing genuine structural homology?

Knowledge of convergence is incomplete unless known by its causes—and its diagnostic signatures.