The gene family responsible for how animals see light has, across hundreds of millions of years, been repeatedly copied, mutated, and put to work doing things that have nothing to do with seeing at all. Deep learning has its own well-worn version of the same basic move: take a network that already learned to do one thing well, and repurpose most of what it learned to do something else entirely, without starting from zero. The comparison between opsin gene evolution and transfer learning is genuinely apt at the level of strategy — both are, at bottom, techniques for reusing an expensive, hard-won piece of machinery for a new job rather than building a new one from scratch. Look closely at how each one actually accomplishes that reuse, though, and a precise, informative difference in mechanism appears.
Scientific Foundation
Opsins are light-sensitive proteins, and the vertebrate visual system as we know it traces back to a single ancestral opsin gene roughly 500 million years ago. Through a sequence of gene duplication events, that single gene gave rise to four spectrally distinct classes of cone opsin, each tuned to a different range of the light spectrum, followed later by the duplication that produced rod opsin from the middle-wave-sensitive cone class. The tuning itself happens through small, specific mutations in the part of the protein that binds its light-absorbing chromophore, shifting exactly which wavelengths a given copy responds to best — a well-studied, ongoing process documented across fish, birds, insects, and primates, including striking independent examples like beetles re-evolving lost blue-light sensitivity through opsin duplication as many as ten separate times.
What makes the comparison to transfer learning genuinely interesting, rather than just a case of “genes get reused,” is that opsins have repeatedly been repurposed for entirely different sensory jobs, not just different colors within vision. Melanopsin, encoded by the OPN4 gene, is expressed in a distinct population of retinal cells and drives circadian photoentrainment, synchronizing the body’s internal clock to the day-night cycle, functioning largely independently of the rod-and-cone system responsible for image-forming sight. Beyond that, researchers have documented opsins recruited into processes with no obvious connection to vision at all: thermosensation, hearing, proprioception — the sense of the body’s own position — and taste, each involving opsin-family proteins doing a genuinely different sensory job than detecting an image.
Cross-Domain Connection
Transfer learning in deep learning solves what looks, on its face, like the same kind of problem. A neural network trained on a large, general task — classifying a huge, broad dataset of images, for instance — develops early layers that extract low-level, highly generic visual features: edges, textures, color gradients, the kind of structure that turns out to be useful for almost any visual task. Standard practice takes advantage of this by freezing those early layers, keeping their learned weights fixed, while replacing and retraining only the later, more task-specific layers on a new, often much smaller target dataset — repurposing a network trained to distinguish a thousand general object categories into one that can, with comparatively little additional data, learn to flag a specific kind of tumor on a medical scan.
What Remains Undemonstrated
Here’s where the mechanisms diverge in a way worth being precise about. Evolution’s strategy for repurposing an opsin is duplicate first, diverge second. The entire gene gets copied, and it’s specifically the existence of that spare, redundant copy that frees one version from the selective pressure protecting the original function — the ancestral copy keeps doing its old job, undisturbed, while the duplicate is free to accumulate mutations and, eventually, drift into a new role entirely, whether that’s a new color of vision or an entirely new sense like thermosensation. The old function is never put at risk, because it’s never actually touched. Transfer learning doesn’t work this way at all. There’s no duplication step. A single copy of the network, already finished being trained on its source task with no ongoing need to keep performing that task in real time, gets modified directly and in place — early layers frozen, later layers overwritten — with nothing analogous to a spare, protected copy continuing to do the original job somewhere else. There’s also no clean biological equivalent of transfer learning’s specific “freeze the first N layers, retrain the rest” rule: cross-modal opsin repurposing accumulates its functional mutations wherever in the duplicated gene they happen to be advantageous, with no equivalent architectural boundary marking which part is “generic” and which part is “task-specific” the way a neural network’s layer structure provides one by design.
Why It Matters
That distinction clarifies something worth taking seriously about why each field’s strategy looks the way it does. Evolution needed a way to explore new functions without ever risking a function currently in active, continuous use, and gene duplication is precisely the solution to that specific constraint — it buys a free, disposable copy to experiment on. Transfer learning doesn’t face that constraint in the same way, because a pretrained network isn’t a living system that needs to keep performing its original task uninterrupted while being repurposed; it’s a finished artifact that can simply be edited. Recognizing that difference matters because it explains why “duplicate then diverge” and “freeze then fine-tune” are both genuinely good answers to the general problem of reuse, without being the same answer — each is well-matched to a different kind of risk its respective field actually has to manage.
Human Dimension
There’s something worth appreciating in the fact that the strategy allowing your eyes to also, distantly and by a circuitous evolutionary path, contribute to how your body senses warmth or keeps time with the sun, isn’t quite the same strategy a machine learning engineer uses to turn an image classifier into a diagnostic tool. Evolution had to be more careful in a specific way computer science doesn’t: it could never risk a currently indispensable function on the chance of discovering a new one, so it learned, over and over, to make a spare copy first and gamble with that instead. The machine learning version of reuse is faster and more direct precisely because it never had that problem to solve in the first place.
Sources:
1. PLOS One — “The Origins of Novel Protein Interactions during Animal Opsin Evolution” — https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0001054
2. PNAS — “Ancestral duplications and highly dynamic opsin gene evolution in percomorph fishes” — https://www.pnas.org/doi/full/10.1073/pnas.1417803112
3. PMC (National Institutes of Health) — “Opsin Gene Duplication in Lepidoptera: Retrotransposition, Sex Linkage, and Gene Expression” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10642689/
4. Scientific Reports (Nature) — “Overcoming the loss of blue sensitivity through opsin duplication in the largest animal group, beetles” — https://www.nature.com/articles/s41598-017-00061-7
5. ScienceDirect — “Evolution of vertebrate visual pigments” — https://www.sciencedirect.com/science/article/pii/S004269890800148X
6. eLife (Reviewed Preprint) — “RHODOPSIN 7: An ancient non-retinal photoreceptor for contrast vision, darkness detection, and circadian regulation” — https://elifesciences.org/reviewed-preprints/109811
7. Eye and Vision (Springer Nature Link) — Buhr, E.D. & Van Gelder, R.N., “Non-visual opsins and their role in circadian photoentrainment” — https://link.springer.com/article/10.1186/s40662-025-00470-0
8. PMC (National Institutes of Health) — “The evolution of irradiance detection: melanopsin and the non-visual opsins” — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2781857/
9. TensorFlow / Keras Documentation — “Transfer learning & fine-tuning” — https://www.tensorflow.org/guide/keras/transfer_learning
10. arXiv — “AutoTune: Automatically Tuning Convolutional Neural Networks for Improved Transfer Learning” — https://arxiv.org/pdf/2005.02165
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 