paper
archived
2022
ⓒ 1
paper:published-28-march-2022

published: 28 March 2022

DOI 10.1016/s2352-4642(22)00126-2OpenAlex W4226055678

TL;DR

Evolutionary transitions in individuality (ETIs) — the process by which lower-level evolutionary units (e.g., unicellular organisms) become components of higher-level individuals (e.g., multicellular organisms) — require interaction structures that compute non-linearly separable functions of component phenotypes, and shallow or single-layer reciprocal networks are formally incapable of this. Watson, Levin, and Buckley introduce the framework of Evolutionary Connectionism, formalizing the equivalence between natural selection acting on heritable variation in inter-unit relationships and unsupervised associative learning in connectionist networks, then extend it to a two-hypothesis account of ETIs as the evolutionary analogue of deep model induction (H1/H2). The core formal result is that only when a developmental process computes a non-linearly separable function — such as XOR or IFF over embryonic particle states — does collective fitness become non-decomposable in the sense required for a genuine new evolutionary unit; monotonic non-linearities, including synergistic but linearly separable functions like AND, leave particle character explanatorily sufficient for particle fitness and thus cannot ground ETIs. Particle plasticity is identified as the concrete individuation mechanism that solves the "heterogeneous functions with homogeneous fitness" (HFHF) problem, and Power's Lotka-Volterra Sudoku model demonstrates that individual-level selection on interaction parameters can produce system-level adaptive organisation without system-level selection. The paper argues this implies that ETIs are predicted by the same three conditions required for deep model induction: a model space permitting asymmetric/deep interaction structures, a distribution of repeated selective perturbations (training samples), and a parsimony pressure functioning as inductive bias — conditions categorically different from genetic relatedness or population bottleneck severity currently emphasised in the ETI literature.

What to take away

  1. 1. Only non-linearly separable collective fitness functions (e.g., XOR, IFF over embryonic particle characters) — not merely monotonic or synergistic non-linearities like AND — make collective character non-decomposable in the sense required for a genuine new evolutionary unit, because only in the non-linearly separable case does the direction of selection on a particle character depend on the context of other particles.
  2. 2. Shallow, single-layer, or fully symmetric reciprocal interaction networks (of the form a↔b, b↔c, c↔a) are formally incapable of computing non-linearly separable functions, just as a single-layer Perceptron cannot compute XOR, which means ETIs require interaction structures with at least one hidden layer of directional asymmetric connections.
  3. 3. The Tudge model demonstrated that natural selection can evolve phenotypic plasticity solving a two-player division-of-labour game (a non-linearly separable function) with homogeneous genotypes, but it assumed genetic relatedness and has not been extended to networks of more than two particles under fully bottom-up selection.
  4. 4. Power's ecological model encoded the constraints of a Sudoku puzzle as competitive Lotka-Volterra interactions among species and showed that individual-level natural selection on interspecific interaction traits — without any system-level selection — caused the community to evolve attractors corresponding to progressively better Sudoku solutions, including puzzles humans find very difficult.
  5. 5. The 'heterogeneous functions with homogeneous fitness' (HFHF) problem specifies that evolutionary individuality requires components to be functionally differentiated (to create non-decomposable collective fitness differences) while having equal reproductive fitness (to prevent particle-level selection from eroding complementarity), and particle plasticity — either phenotypic or reproductive — is the mechanism that decouples function from fitness.
  6. 6. Evolutionary Connectionism formalizes the equivalence between random variation and selection acting on heritable variation in pairwise inter-unit relationships and unsupervised Hebbian associative learning, so that an ecological community evolving interaction strengths is formally equivalent to a recurrent neural network learning an associative memory without a system-level reward signal.
  7. 7. The paper raises the open hypothesis (H2) that the three conditions sufficient for deep model induction in machine learning — a model space capable of representing non-linearly separable functions, a representative distribution of training samples (repeated selective perturbations/shocks), and a parsimony pressure as inductive bias — are jointly predictive of the conditions under which bottom-up natural selection can produce an ETI.
  8. 8. Basal cognition — substrate-independent information integration and collective action observed in non-neural developmental systems such as planarian regeneration and bioelectric signalling, as developed by Levin and colleagues — is identified as the organismic-level process that implements the computation of non-decomposable collective characters, directly linking organismic individuality to the formal requirements for evolutionary individuality.
  9. 9. A researcher wishing to test H1 empirically or in simulation should operationalise the transition criterion as follows: construct collectives in which the developmental interaction structure is progressively made asymmetric and directional (rather than symmetric and reciprocal), measure whether collective fitness becomes non-linearly separable as a function of embryonic particle states, and track whether particle-level character ceases to be a reliable predictor of particle fitness across contexts.
  10. 10. The paper explicitly acknowledges that it has not yet produced a unified mathematical model linking non-decomposable collective characters, the collective-level response to selection (Bourrat's non-aggregative Price component), and the direction of selection on plasticity parameters, and predicts that this gradient will prove equivalent to the learning gradient in a connectionist system applied to a non-linearly separable objective function.

Peer brief — for seminar discussion

Watson, Levin, and Buckley (Frontiers in Ecology and Evolution, 10:823588, 2022) address the foundational chicken-and-egg problem of evolutionary transitions in individuality (ETIs): how can bottom-up selection on existing lower-level units produce the organised interaction structures necessary to constitute a new higher-level evolutionary unit, without presupposing that unit? The paper introduces Evolutionary Connectionism as its core theoretical framework, which formalizes the functional equivalence between natural selection acting on heritable variation in inter-unit relationships and unsupervised associative learning in connectionist neural networks, then extends this equivalence into two explicit hypotheses (H1 and H2) about what kinds of interaction structures are necessary for ETIs and under what conditions they can evolve. The load-bearing finding is the formal identification of non-linearly separable functions — the class of functions exemplified by XOR and IFF, in which the sign of the effect of one input on the output reverses depending on the value of another input — as the precise criterion for non-decomposable collective characters. This is a strictly stronger condition than prior accounts of non-aggregative interactions (Bourrat 2021), which included monotonic non-linearities: a synergistic function like AND is non-aggregative but remains linearly separable, meaning that the direction of selection on a particle character is still consistent across contexts, leaving the collective explanatorily redundant with respect to particle fitness. Only when the collective phenotype is a non-linearly separable function of embryonic particle states does collective character, rather than particle character, determine the direction of selection on particles. The paper further argues that shallow or fully symmetric reciprocal interaction architectures — the kind assumed in prior evolutionary connectionism models of Evo-Eco — cannot compute non-linearly separable functions (exactly as a single-layer Perceptron cannot compute XOR), so depth in the interaction structure is a necessary condition for an ETI. Particle plasticity is identified as the individuation mechanism that solves the HFHF problem by decoupling function from fitness, and Power's Lotka-Volterra Sudoku model (Power 2019, University of Southampton PhD) is cited as the clearest demonstration that individual-level selection on interaction parameters can produce system-level adaptive organisation without system-level selection. H2 predicts that ETIs require exactly the three conditions for deep model induction: a model space including asymmetric directional interaction structures with hidden variables; repeated perturbations providing a distribution of selective environments analogous to a training set; and a parsimony pressure serving as inductive bias. These predictions are categorically different from the variables currently foregrounded in ETI research — genetic relatedness, severity of developmental bottlenecks, population structure — which the paper treats as neither necessary nor sufficient in isolation. An alternative theoretical approach the paper could have used is ecological scaffolding (Black, Bourrat, and Rainey 2020), which solves the same chicken-and-egg problem by invoking exogenous ecological conditions that temporarily impose population structure; the paper explicitly contrasts its approach with scaffolding, arguing that scaffolding produces canalisation of fortuitous states whereas evolutionary connectionism produces genuine adaptive generalisation to novel high-fitness configurations. The most contestable aspect is the inferential gap between the formal analogy and causal sufficiency: the paper demonstrates that ETI interaction structures are formally isomorphic to deep model induction, and that individual-level selection can produce system-level organisation in shallow models (Evo-Eco) and deep organisation under system-level selection (Evo-Devo), but no single model yet demonstrates deep interaction structures evolving under purely bottom-up selection and crossing an ETI threshold. The paper acknowledges this explicitly — listing four component results that each capture part of H1 but none the whole — which means H2 remains predictive rather than demonstrated. A critical reader would push back on whether the formal equivalence between Hebbian relaxation of connections and natural selection on second-order traits holds in sufficiently general biological parameter regimes, and whether parsimony pressure, which the paper treats as a mild and ubiquitous assumption, is actually identifiable as a distinct selective mechanism in specific biological ETI substrates such as the origin of chromosomes or eukaryotic organelles.

Findings (1)

Claims (3)

Hypotheses (1)

Original abstract (expand)

Mensah AM, Campbell H, Stowe S, et al. Risk of SARS-CoV-2 reinfections in children: a prospective national surveillance study between January, 2020, and July, 2021, in England. Lancet Child Adolesc Health 2022; published online March 28. https://doi.org/10.1016/S2352-4642(22)00059-1. We now offer open access for authors based at JISC-participating UK institutions for papers accepted after Jan 1, 2022. As a result, the copyright line for this Article has been updated to Crown Copyright © 2022 Published by Elsevier Ltd.

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

Similar preprints — Semantic Scholar

Cross-corpus bridges (12)

same_concept_as · Nomic cosine

External markdown files that talk about the same concept as this entity.

  • aboutblank_kb
    Evolutionary Transitions In Individuality (Etis)frameworks/evolutionary-transitions-in-individuality-theory.md0.870
  • aboutblank_kb
    Evolutionary Transitions In Individuality (Etis)concepts/biology/evolutionary-transitions-in-individuality-etis.md0.869
  • aboutblank_kb
    Connectionist Approach To Evolutionary Transitions In Individualityframeworks/connectionist-approach-to-evolutionary-transitions-in-individuality.md0.833
  • aboutblank_kb
    How do non-linearly separable interactions between components enable the emergence of new levels of biological individuality?questions/how-do-nonlinearly-separable-interactions-between-components-enable.md0.822
  • aboutblank_kb
    Connectionist Framework Of Individualityframeworks/connectionist-framework-of-individuality.md0.818
  • aboutblank_kb
    Design For Individualityframeworks/design-for-individuality.md0.818
  • aboutblank_kb
    How do evolutionary transitions in individuality emerge from lower-level coordination?questions/how-do-evolutionary-transitions-in-individuality-emerge-from.md0.817
  • aboutblank_kb
    How do evolutionary transitions in individuality relate to bioelectric integration?questions/how-do-evolutionary-transitions-in-individuality-relate-to.md0.816
  • aboutblank_kb
    How can traits that increase non-aggregative components of collective heritability evolve through bottom-up selection when they may ultimately oppose particle-level fitness?questions/how-can-traits-that-increase-nonaggregative-components-of.md0.816
  • aboutblank_kb
    How are the conditions for deep model induction in connectionist learning systems predictive of ETI evolution conditions?questions/how-are-the-conditions-for-deep-model-induction.md0.815
  • aboutblank_kb
    Under what conditions can deep interaction structures computing non-linearly separable functions evolve through bottom-up selection?questions/under-what-conditions-can-deep-interaction-structures-computing.md0.815
  • aboutblank_kb
    Evolutionary Connectionismframeworks/evolutionary-connectionism.md0.813