Evolution May Be a Form of Multilevel Learning

The Surprising Link Between Evolution and Learning

Imagine watching a slime mold solve a maze. It doesn't have a brain. It doesn't have neurons. Yet it finds the shortest path to food, remembers where it has been, and even anticipates periodic events. For decades, biologists called this behavior "learning" in scare quotes, as if the organism was faking it.

Now a team of physicists and biologists has proposed something far stranger. Evolution itself, they argue, might be a form of learning. Not metaphorically. Mathematically.

In a 2022 paper published in the Proceedings of the National Academy of Sciences, Vitaly Vanchurin of the University of Minnesota, along with Yuri Wolf, Mikhail Katsnelson, and Eugene Koonin, laid out a formal correspondence between evolution, learning algorithms, and renormalizable physical theories (Vanchurin et al., 2022). The authors showed that the core components of evolution natural selection, replication, mutation can be derived from the same equations that describe how neural networks learn. The implication is that evolution and learning are not just analogous processes. They might be the same process, running at different scales.

This is not a gentle tweak to evolutionary theory. It is a proposal to rewrite the mathematical language of biology.

What Does It Mean for Evolution to "Learn"?

The standard view of evolution

The textbook version of evolution is deceptively simple. Organisms vary. Some variations help survival and reproduction. Those traits get passed down. Over generations, populations change. This is the engine of adaptation.

But the textbook version struggles with one big question: How do new levels of complexity emerge? How do single cells become multicellular organisms? How do organisms form societies? How do societies develop language and culture? Standard evolutionary theory describes how existing traits get refined, but it has no good mathematical framework for explaining how entirely new levels of organization arise.

The learning analogy

Vanchurin and his colleagues started with a simple observation. When a neural network learns, it adjusts its internal parameters to minimize error. When a population evolves, it adjusts its genetic composition to maximize fitness. Both processes involve variation, selection, and retention. Both operate over time. Both produce systems that appear to be solving problems.

The authors formalized this intuition. They showed that the mathematical structure of learning in neural networks the process of adjusting weights to reduce prediction error maps directly onto the dynamics of evolving populations (Vanchurin et al., 2022). Replication, mutation, and natural selection emerge naturally from the learning equations. They are not add-ons. They are necessary consequences.

This is the first major claim of the paper: Evolution can be understood as a learning algorithm that operates on genetic material rather than on synaptic weights.

The Hidden Architecture: Renormalization and Multilevel Learning

What is renormalization?

Renormalization is a concept from physics that sounds arcane but is actually intuitive. It is a method for describing a system at different scales. When you zoom out from atoms to molecules to cells to tissues, the rules that govern behavior change. Renormalization provides a mathematical way to connect these levels.

Vanchurin and his colleagues realized that evolution faces the same problem. Genes operate at one scale. Organisms operate at another. Ecosystems at yet another. The rules of evolution at each level are not independent. They are linked by the same kind of mathematical transformations that physicists use to connect quantum mechanics to thermodynamics.

Learning at multiple levels

Here is where the paper gets genuinely radical. The authors argue that learning does not happen only at the genetic level. It happens at every level of biological organization simultaneously (Vanchurin et al., 2022).

Consider a flock of birds. Each bird follows simple rules: stay close to neighbors, avoid collisions, match speed. But the flock as a whole learns to navigate, to find food, to avoid predators. The flock's behavior is not programmed into any individual bird. It emerges from the collective learning dynamics of the group.

The same logic applies to cells in a body, to neurons in a brain, to individuals in a society. At each level, there is a learning process: variation, selection, retention. And these processes are not separate. They are coupled. Changes at the genetic level constrain what is possible at the organismal level. Changes at the organismal level feed back to shape what genes are selected.

This is what the authors call "multilevel learning." Evolution is not a single process running on genes. It is a hierarchy of learning processes, each operating at a different scale, each influencing the others.

How the Study Was Done: Building the Mathematical Bridge

The method

Vanchurin and his colleagues did not run experiments on fruit flies or sequence genomes. They built mathematical models. Specifically, they constructed a formal correspondence between three domains:

• ▸Evolutionary dynamics: The equations that describe how allele frequencies change over time, how populations evolve, how new species arise.

• ▸Learning theory: The mathematics of neural networks, reinforcement learning, and Bayesian inference systems that improve their performance through experience.

• ▸Renormalization group theory: The physics framework for describing systems at multiple scales.

The authors showed that the key equations from each domain share the same mathematical structure. The Fisher equation from population genetics, which describes how genetic variation spreads, is mathematically equivalent to the learning rule in certain neural networks. The Price equation, which partitions evolutionary change into components, maps onto the backpropagation algorithm used to train deep learning systems.

What they found

The correspondence was not approximate. It was exact. Vanchurin and his colleagues demonstrated that given the right assumptions, the dynamics of evolution and the dynamics of learning are governed by the same underlying equations (Vanchurin et al., 2022).

This means that concepts from learning theory can be imported directly into evolutionary biology. For example:

• ▸Overfitting in machine learning occurs when a model learns noise instead of signal. The authors suggest that the same phenomenon occurs in evolution when a population becomes too specialized to a narrow environment and loses the ability to adapt to change.

• ▸Regularization techniques that prevent overfitting in neural networks have evolutionary analogues: mechanisms that maintain genetic diversity or slow down adaptation to preserve flexibility.

• ▸Transfer learning where knowledge gained in one task helps learning another maps onto the evolution of modular genetic networks that can be reused across different contexts.

The authors did not just claim these analogies. They demonstrated the formal equivalence.

What This Actually Means

A new mathematical language for biology

The most immediate implication is practical. Evolutionary biology has long been a descriptive science. It tells stories about how traits evolved. But it has struggled to build predictive mathematical models that work across scales. The Vanchurin et al. framework provides a unified mathematical language.

If evolution is a learning algorithm, then the entire toolkit of machine learning becomes available to biologists. Algorithms for optimization, for generalization, for hierarchical learning all of them can be applied to evolutionary questions. This is not about using AI to simulate evolution. It is about recognizing that evolution and AI are running the same algorithms.

A resolution to the levels problem

The multilevel learning framework offers a clean answer to the puzzle of evolutionary transitions. Why did life go from single cells to multicellular organisms? Because learning at the cellular level created a new level of organization that could itself learn. The transition was not a lucky accident. It was a consequence of the learning dynamics themselves.

When a system learns at one level, it naturally generates higher level structures that can also learn. This is not speculation. It is what happens in deep neural networks, where lower layers learn simple features and higher layers learn complex combinations. The authors argue that the same hierarchical learning occurs in biological evolution.

A bridge between biology and physics

The renormalization connection is the deepest part of the paper. It suggests that the multilevel structure of evolution is not a biological accident. It is a reflection of a deeper principle about how complex systems organize themselves.

In physics, renormalization explains why the same laws of thermodynamics apply to a gas in a box and to the universe as a whole. The details of individual molecules wash out, and the large scale behavior becomes simple. Vanchurin and his colleagues propose that the same thing happens in evolution. The messy details of individual mutations and selective events wash out, and the large scale behavior of evolving systems follows simple learning dynamics.

This does not mean that evolution is deterministic. It means that the statistical patterns of evolution are governed by the same mathematics as learning.

What This Does Not Prove

It is not a claim about consciousness

Let me be clear about what the authors are not saying. They are not claiming that bacteria are conscious. They are not claiming that evolution has a goal or a purpose. Learning in the mathematical sense does not require awareness. A neural network learns without knowing it is learning. An evolving population learns without knowing it is evolving. The word "learning" here refers to a formal process of parameter adjustment, not to subjective experience.

It is not a replacement for Darwinian evolution

The multilevel learning framework does not contradict natural selection. It subsumes it. Natural selection emerges from the learning dynamics at the genetic level. The authors are not proposing an alternative to Darwin. They are proposing a deeper mathematical structure that explains why Darwinian evolution works.

It is an open question whether the analogy is exact

The formal correspondence that Vanchurin et al. demonstrated holds under specific mathematical assumptions. Real biological systems are messier. Populations are finite. Environments change. Mutations are not always random. The authors acknowledge that the analogy is not perfect. The question is whether the imperfections are noise or signal. Do the deviations from the learning framework reveal something important about biology, or are they just details that will be absorbed as the theory develops?

This is not a weakness of the paper. It is the most interesting open question it raises.

The Bigger Picture: Why This Changes Everything

Evolution as a universal learning process

If Vanchurin and his colleagues are right, then evolution is not just a biological phenomenon. It is a specific instance of a universal learning process that operates wherever there is variation, selection, and retention. This includes genetic evolution, but also immune systems, neural development, cultural change, and even the evolution of ideas.

The authors are not the first to notice these analogies. But they are the first to provide a rigorous mathematical foundation for them. This moves the conversation from "isn't it interesting that evolution and learning look similar" to "here is the exact equation that governs both."

Implications for artificial intelligence

The multilevel learning framework has a practical consequence for AI research. If evolution is a form of learning, then the algorithms that work in machine learning might also work for evolutionary optimization. And conversely, the strategies that evolution has discovered over billions of years might improve AI systems.

Consider the problem of catastrophic forgetting in neural networks. When a network learns a new task, it often forgets how to do old tasks. Evolution solved this problem long ago. Organisms retain the ability to perform ancestral functions even as they adapt to new environments. The mechanisms evolution uses modularity, redundancy, regularization could be directly applied to improve AI.

A new research program

The Vanchurin et al. paper is not a finished theory. It is a proposal for a new research program. The authors lay out several directions for future work:

• ▸Testing whether the learning dynamics framework can predict evolutionary outcomes in laboratory populations.

• ▸Searching for signatures of multilevel learning in the fossil record, such as periods of rapid innovation followed by stabilization.

• ▸Developing mathematical tools to identify the relevant levels of learning in complex biological systems.

• ▸Applying the framework to understand the evolution of cancer, where cell level learning (proliferation) conflicts with organism level learning (homeostasis).

What This Actually Means (The Takeaway)

• ▸Evolution and learning share the same mathematics. The equations that describe how neural networks learn are the same equations that describe how populations evolve. This is not metaphor. It is formal equivalence.

• ▸Biological complexity emerges from multilevel learning. New levels of organization cells, organisms, societies arise because learning at one level naturally creates conditions for learning at higher levels. Evolutionary transitions are not accidents. They are consequences.

• ▸The renormalization group provides a bridge between biology and physics. The same mathematical framework that connects quantum mechanics to thermodynamics also connects genetic evolution to ecosystem dynamics. This suggests deep principles about how complex systems organize across scales.

• ▸Evolutionary biology gains a new toolkit. Concepts from machine learning overfitting, regularization, transfer learning become directly applicable to evolutionary questions. This is a practical tool, not just a philosophical insight.

• ▸The theory makes testable predictions. If evolution is a learning algorithm, then populations should exhibit learning phenomena like generalization, memorization, and catastrophic forgetting. These predictions can be tested in laboratory evolution experiments.

The Vanchurin et al. framework is one of those rare papers that changes how you see the world. After reading it, you cannot unsee the learning dynamics everywhere. The flock of birds is not just moving. It is solving a problem. The immune system is not just fighting infection. It is updating its model of the world. The tree in your backyard is not just growing. It is learning.

And evolution is not just a historical process that shaped life on Earth. It is a universal algorithm that runs wherever there is variation, selection, and time. The same algorithm that built your brain is running inside your brain. The learner is learning itself.