Despite its huge reach, the theory of evolution is simple. It rests on three pillars: variation, selection and inheritance. Variation stems from random genetic mutations, which create genetically distinct individuals. Natural selection favours “fitter” individuals, with those better suited than others to a particular environment prospering and producing the most offspring. Inheritance means that these well-adapted individuals pass their characteristics down the generations. All this eventually leads to new adaptations and new species.
At first glance, there is no need for learning in this process. In fact, to invoke it at all risks violating one of evolution’s most important principles. When we learn, we in some way anticipate the future, combining solutions from past experience with knowledge of present conditions to develop a strategy for what we think will come next. But evolution can’t see the future: its exploration is born out of random mutations selected or rejected by current circumstances, so it is blind to the challenges to come.
But then again, learning organisms can’t actually see the future. When we cross a road, we can’t anticipate all traffic movements, but we have a memory bank of solutions that have worked before. We develop a strategy based on those – and if it proves successful, we call on that newly learned experience next time. That’s not too dissimilar to what natural selection does when it reuses successful variants from the past, such as the flowers of bee orchids that are unusually good at attracting bees, or the mouthparts of mosquitoes that work like hypodermic syringes and are particularly effective at sucking blood.
Some now think the similarities between learning and evolution go more than skin-deep – and that our understanding of one could help understand the other. Since the early days of computer science, researchers have been developing algorithms – iterative rules – that allow computers to combine banked knowledge with fresh information to create new outputs, and so mimic processes involved in learning and intelligence. In recent years, such learning algorithms have come to underlie much technology that we take for granted, from Google searches to credit-scoring systems. Could that well-stocked toolkit now prise open the secrets of evolution? “The analogy between evolution and learning has been around for a long time,” says Richard Watson of the University of Southampton, UK. “But the thing that’s new is the idea of using learning theory to radically expand our understanding of how evolution works.”
A pioneer of this approach is Leslie Valiant, a computational theorist at Harvard University. In his 2013 book Probably Approximately Correct, he described how the workings of evolution equate to a relatively simple learning algorithm known as Bayesian updating. Used to model everything from celestial mechanics to human decision-making computationally, this type of learning entails starting with many hypotheses and pinpointing the best ones using new information as it becomes available. Replace the hypotheses you want to test with the organisms in a population, Valiant showed, and natural selection amounts to incorporating new information from the surrounding environment to home in on the best-adapted organisms.
A learning network
That could be just coincidence. But in 2014, Erick Chastain at Rutgers University in New Brunswick, New Jersey, and his colleagues found a similar equivalence between evolution in a sexually reproducing population and another learning model called the multiplicative weights update algorithm. This presumes there may be many potential solutions to a problem, and the key to finding the best lies in weighting their promise on the basis of past performance. Applying this algorithm and assuming that natural selection gives more weight to previously successful solutions was enough to reproduce how, over generations, evolution homes in on the gene variants with the highest overall fitness.
Such parallels left Watson wondering how a model that more closely follows the genetic changes underpinning evolution might look. Not so long ago, we naively talked about genes “for” particular traits, and assumed for example that humans, being so complex, would have lots of genes. When in the 1990s two groups were vying to sequence the human genome, they believed they would identify some 100,000 genes. To everyone’s surprise, they discovered we have fewer than 25,000. The reason, we now know, is that genes are team players: their activity is regulated by other genes, creating a network of connections. The whole is thus capable of much more than the sum of its parts.
These connections mean that mutations, whether caused by spontaneous chemical changes or faulty DNA repair processes, don’t just alter single genes. When a mutation changes one gene, the activity of many others in the network can change in concert. The network’s organisation is itself a product of past evolution, because natural selection rewards gene associations that increase fitness. This allows your genotype (the set of genes you inherit from your parents) to solve the problem of creating a well-adapted phenotype (the set of outward characteristics that adds up to you). “In evolution, the problem is to produce a phenotype that is fit in a given environment, and the way to do it is to make connections between genes – to learn what goes together,” says Watson.
Watson’s insight was to realize that this whole process has a lot in common with the workings of one of the cleverest learners we know – the human brain. Our brains consist of neurons connected via synapses. Connections between two neurons are strengthened when they are activated at the same time or by the same stimulus, a phenomenon encapsulated by the phrase “neurons that fire together wire together”. When we learn, we alter the strengths of connections, making networks of associations capable of problem-solving.
This is called Hebbian learning after neuropsychologist Donald Hebb, who first described it in the mid-20th century. Simple models based on these networks can do surprisingly clever things, such as recognising and classifying objects, generalising behaviour from examples, and learning to solve optimisation problems. If evolution works in equivalent ways, Watson realised, that could explain why it is such a good problem-solver, creating all that complexity in such short order.
Spontaneous solutions
Working with Günter Wagner from Yale University and others, Watson built a model network in which genes can either increase or reduce each other’s activity, as they do in nature. Each network configuration controls how the genes within it interact to give rise to a different phenotype, presented in the form of a pixelated image on a screen. The modellers evolved the network by randomly changing the gene connections, one mutation at a time, and selecting those networks that produced an image with a closer resemblance to one deemed to be the optimal phenotype – a picture of Darwin’s face. Thus guided, the evolving system eventually reproduced this image, at which point the team used the same process to teach it to reproduce Hebb’s face.
But here came the surprise. The modellers then removed the selection pressure guiding the system towards mugshots of Darwin or Hebb. Any old mutation that arose was allowed to survive. But the system did not produce a random image, or a Darwin-Hebb mash-up. Instead, it produced one or the other face – and as little as a single gene mutation was enough to trigger a flip between the two. In other words, a model that simply took account of genes’ networked nature showed that when the genotype had learned solutions, it could remember them and reproduce them in different environments – as indeed our brains can.
Evidence for learning in this sense is often seen in the natural world, for instance in the way a crocodile genome can produce a male or female crocodile depending on the temperature at which the egg is incubated. But learning the way our brains do it is not just about remembering and reproducing past solutions. “A real learning system also has to be able to generalise – to produce good solutions even in new situations it hasn’t encountered before,” says Watson. Think crossing a road you’ve never crossed before versus crossing a familiar one.
This generalization ability rests in recognizing similarities between new and old problems, so as to combine the building blocks of past solutions to tackle the problem at hand. And as another model created by Watson and his colleagues showed last year, this kind of learning is also what a gene network does under the pressure of natural selection. The cost associated with making gene connections – proteins must be produced and energy expended – favours networks with fewer connections. Subsets of connections that work well together become bound tightly in blocks that themselves are only loosely associated. Just as our brains do, natural selection memorises partial solutions – and these building blocks are embedded in the structure of the gene network (arxiv.org/abs/1508.06854).
This way of working allows genotypes to generate phenotypes that are both complex and flexible. “If past selection has shaped the building blocks well, it can make solving new problems look easy,” says Watson. Instead of merely making limbs longer or shorter, for example, evolution can change whether forelimbs and hindlimbs evolve independently or together. A single mutation that changes connections in the network can lengthen all four legs of a giraffe, or allow a bat to increase its wingspan without getting too leggy. And a feather or an eye needn’t be generated from scratch, but can evolve by mixing and matching building blocks that have served well in the past (see diagram [attached in post]).
This ability to learn needs no supernatural intervention – it is an inevitable product of random variation and selection acting on gene networks. “Far from being blind or dumb, evolution is very smart,” says Watson.
Watson’s idea has caught the attention of respected evolutionary theorists, among them Eörs Szathmáry of the Parmenides Foundation in Munich, Germany. “It is absolutely new,” he says. “I thought that the idea was so fascinating and so interesting that I should put some support behind it.” Earlier this year, he and Watson collaborated on a paper called “How Can Evolution Learn?” to discuss some of its implications (Trends in Ecology and Evolution, vol 31, p 146).
“Evolution looking like the product of intelligence is exactly what you’d expect“
For a start, if evolution learns, by definition it must get better at what it does. It will not only evolve new adaptations, but improve its ability to do so. This notion, known as the evolution of evolvability, has been around for some time, but is contentious because it seems to require forethought. No longer. “If you can do learning, then you are able to generalise from past experience and generate potentially useful and novel combinations,” says Szathmáry. “Then you can get evolvability.”
Applying similar ideas might also begin to explain how ecosystems evolve (see “Eco-learning“). More speculatively, Watson and Szathmáry suggest that the marriage between learning theory and evolutionary theory could throw light on the giant leaps made by evolution in the past 3.8 billion years. These “major transitions”, an idea first formulated by Szathmáry and John Maynard Smith in the 1990s, include the jumps from replicating molecules to cellular organisms, from single-celled to multicellular organisms and from asexual to sexual reproduction. Szathmáry and Watson think the key might lie in a model known as deep learning.
This was how Google DeepMind beat the world’s top player at the ancient and fiendish game of Go earlier this month. It is based on Hebbian learning, with the difference that it “freezes” successive levels of a network once it has learned as much as it can, using the information acquired as the starting point for the next level. “It’s intriguing that evolutionary algorithms exploiting deep learning can solve problems that single-level evolution cannot,” says Watson – although he admits the details of the parallel are still to be worked out. If we could tease out the circumstances required to produce a major transition, that might suggest where evolution is heading next – or even how to engineer a transition. For example, says Watson, it might show us how to transform a community of microbes into a true multicellular organism.
Other evolution researchers are also intrigued. “Watson and Szathmáry are right in recognising that a species’ evolutionary history structures its genes in much the same way that an individual’s learning history structures its mind,” says Mark Pagel of the University of Reading, UK. David Sloan Wilson at Binghamton University, New York, thinks it could be an important step forward too. “In the past, it has been heretical to think about evolution as a forward-looking process, but the analogy with learning – itself a product of evolution – is quite plausible,” he says.
Szathmáry thinks we can fruitfully see that analogy from both ends. If evolution and cognitive learning are based on the same principles, we can use our understanding of either to throw new light on the other. With that in mind, he is now co-opting evolutionary theory to investigate the long-standing puzzle of how infants learn language so easily with no formal teaching and little other input.
Those infants may now grow up with a better grasp on the processes underlying that greatest of theories, evolution by natural selection. If evolution looks smart, that’s because it is, says Watson. “The observation that evolutionary adaptations look like the product of intelligence isn’t evidence against Darwinian evolution – it’s exactly what you should expect.”
Eco-learning
How does a forest develop the ability to share limited resources such as light and water? It’s a puzzle how such seemingly harmonious environments develop from a collection of individuals of different species. “There’s currently no general theory of how ecosystems evolve,” says Richard Watson at the University of Southampton, UK.
By thinking about the genes within organisms as networks, Watson has shown that evolution can learn from the past to solve new problems of survival (see main story). But the same might apply to ecological networks, he thinks.
There is a big difference: ecosystems must “learn” unaided by natural selection because it only favours fit individuals, not fit communities. Computer scientists have a way to model such “unsupervised” learning. Instead of guiding an evolving network towards a solution to a given problem, they use a rule that reinforces already common correlations. Such algorithms are known to efficiently discover categories, clusters and regularities in data sets. Working with his Southampton colleague Daniel Power and others, Watson has used them to model learning in the evolution of an ecological network (Biology Direct, vol 10, p 69).
Remarkably, ecosystem networks allowed to evolve in this way retain collective memories – information about ecological interactions that have produced successful structures and behaviours in the past. This could explain why a damaged coral reef can revert to its former composition if left to recover, or why a rainforest can recover from fragmentation.
It also suggests that if an ecosystem faces a challenge it has encountered in the past, it may be more able to recover than in the face of a new challenge. In theory, an ecosystem’s collective memory could be extremely long, because of being etched into the genomes of the organisms that comprise it. An ecosystem that has experienced warming in the past, for example, might be better equipped to cope with global warming now.
