How evolution learns to generalise: Using the principles of learning theory to understand the evolution of developmental organisation

Abstract

One of the most intriguing questions in evolution is how organisms exhibit suitable phenotypic variation to rapidly adapt in novel selective environments. Such variability is crucial for evolvability, but poorly understood. In particular, how can natural selection favour developmental organisations that facilitate adaptive evolution in previously unseen environments? Such a capacity suggests foresight that is incompatible with the short-sighted concept of natural selection. A potential resolution is provided by the idea that evolution may discover and exploit information not only about the particular phenotypes selected in the past, but their underlying structural regularities: new phenotypes, with the same underlying regularities, but novel particulars, may then be useful in new environments. If true, we still need to understand the conditions in which natural selection will discover such deep regularities rather than exploiting ‘quick fixes’ (i.e., fixes that provide adaptive phenotypes in the short term, but limit future evolvability). Here we argue that the ability of evolution to discover such regularities is formally analogous to learning principles, familiar in humans and machines, that enable generalisation from past experience. Conversely, natural selection that fails to enhance evolvability is directly analogous to the learning problem of over-fitting and the subsequent failure to generalise. We support the conclusion that evolving systems and learning systems are different instantiations of the same algorithmic principles by showing that existing results from the learning domain can be transferred to the evolution domain. Specifically, we show that conditions that alleviate over-fitting in learning systems successfully predict which biological conditions (e.g., environmental variation, regularity, noise or a pressure for developmental simplicity) enhance evolvability. This equivalence provides access to a well-developed theoretical framework from learning theory that enables a characterisation of the general conditions for the evolution of evolvability.

How Can Evolution Learn?

“…Example prediction: Short-term selection can increase long-term evolvability if it benefits from an appropriate inductive bias, for example, that the genotype-phenotype map is complex enough to represent structure (epistatic interactions) in the selective environment but simple enough to avoid overfitting that structure [51,53] 2. Can ecosystem functions be adapted without ecosystem selection?…”

“…Accordingly, the possibility that evolution can learn from experience to favourably bias future exploration need not be any more mysterious than the basic result that learning from a training set can produce good generalisation on an unseen test set [51]. This also sheds light on the tension between robustness and evolvability.…”

“…To avoid this, the complexity of a model can be limited (e.g., by applying a parsimony pressure) to prevent memorisation of unnecessary details and force solutions to capture deeper regularities (Box 1). This explains why a cost of connections increases evolved modularity and improves evolvability [51,53]. discover categories, clusters, or regularities inherent in the training samples and hence reduce the effective dimensionality of the data [14] (Box 2).…”

Morphological Modularity Can Enable the Evolution of Robot Behavior to Scale Linearly with the Number of Environmental Features

“…Catastrophic forgetting (French, 1999) is a common problem in machine learning, whereby a learner must forget something in order to learn something new. Third, complex predictive models and dense, non-modular networks can suffer from the pathology of overfitting: they fail to generalize to novel environments (Kouvaris et al, 2015). Modular networks can avoid overfitting by internally reflecting the modularity in its environment: it responds appropriately in a "new" environment, which is actually just an unfamiliar combination of familiar percepts.…”

How to fit in: The learning principles of cell differentiation

“…random variation and selection) on gene regulatory networks (GRNs) underlying cell plasticity and cell differentiation can exhibit comparable learning capabilities is a non-trivial question. Although during the last years it has been shown that learning principles operate across a plethora of biological phenomena [21]- [26], it remains unknown if these principles apply to plastic cells performing categorisation learning tasks. Furthermore, most studies of evolutionary learning rely on simple NN-like modelling strategies, employing simplifying assumptions that are common in models of artificial neural networks but not appropriate for natural gene networks [17], [21], [26].…”