Table 1 A unified view of evolutionary and learning processes, as part of a theory of high-dimensional adaptations.

1. Learning in evolution

How does a Darwinian process over populations of non-neural systems give rise to algorithmic elements of learning?

Emergent neural network-like dynamics. Connection strengths between units adapt and store information about the environment; activation dynamics of units, parametrized by the adapted interaction strengths, compute⁶. Paradigmatic examples include Hebbian learning in gene regulatory networks⁷ and in ecosystems⁸. Computation, on a developmental and ecological timescale, gives rise to autoassociative dynamics. As a result, developmental systems and ecosystems, according to this theory, store memory traces about past environments, making the system easily re-adapt to those environments once they recur. Note that local Hebbian dynamics is either a consequence of direct selection for efficiency on a population of networks, as in gene regulatory networks, or it is not selected for, as in ecosystems⁹

Emergent Bayesian dynamics. Type frequencies store and update information about the environment, in a way that it is isomorphic to the competition dynamics of statistical hypotheses in a Bayesian setting^10,11,12. Structural-dynamical equivalences extend far beyond those of replicator dynamics and Bayesian update¹³. They include a mapping between multilevel selection and Bayesian inference in hierarchical models¹⁴, between quasispecies dynamics and filtering in hidden Markov models¹⁵, and between evolutionary-ecological competition dynamics and expectation-maximization optimization of mixture models¹³. In short, Darwinian dynamics can be construed as accumulating evidence for models of the environment that are entailed in the phenotype^13,16,17. This provides an interesting perspective that connects the normative formulation of natural selection in the brain to Bayesian model selection (i.e., structure learning) under the Bayesian brain hypothesis^{14,18,19,20,21}

Emergent sampling algorithms implemented by the evolution of finite populations¹². Sampling non-trivial probability distributions is considered to be a more and more fundamental module of mental computations²². Adaptation to appropriate fitness landscapes by a finite population results in sampling a corresponding distribution, making use of the stochasticity provided by genetic drift. Two examples are the evolution of types in finite population models with a regime of strong selection and weak mutation giving rise to Markov Chain Monte Carlo dynamics²³ and the equivalence of the evolution of relative frequencies in the Wright-Fisher model and a fundamental approximative statistical inference algorithm, particle filtering¹²

Emergence of evolutionary novelties as insights in the course of problem solving. At a macroevolutionary timescale, evolution proceeds by long nearly-static periods punctuated by sudden “inventions” of novel phenotypic solutions; this has been conceptually and algoritmically linked to search for out-of-the-box solutions in insight problems, where a sudden conscious emergence of a right solution follows a long incubation period marked by “stasis in solution space”^24,25

2. Evolution in learning

How does a Darwinian process emerge from local learning rules?

The low-level channel of information passing between the system and the environment is any mechanism that provide plasticity and adaptation at a lower algorithmic level; heritable variation and selection is a higher-level emergent property. Darwinian neurodynamics exemplifies this approach. As opposed to proposed mechanisms under the umbrella conventionally called neural Darwinism^26,27, Darwinian neurodynamics (DN) decouples the concept of population in the neural and evolutionary sense: the replicators are neural activity patterns, not anatomical structures⁴. In particular, a network composed of the same exact neurons and synapses might produce different activity patterns at different time instances and therefore give rise to different replicators in the evolutionary sense. Another fundamental difference between neural Darwinism and Darwinian neurodynamics is that the latter performs bona fide evolutionary search, in which multiple rounds of selection acts on heritable variation