The collective intelligence of evolution and development

What to take away

1. 1. Evolutionary connectionism — the paper's central framework — holds that random variation and selection acting on interaction networks is formally equivalent to connectionist learning, enabling the transfer of the entire machine-learning toolset into evolutionary and developmental theory.
2. 2. Ectopic eyes transplanted to the tails of Xenopus tadpoles still confer light-mediated learning because the eye primordium cells successfully wire optic nerves to the spinal cord rather than the brain, demonstrating morphogenetic problem-solving in anatomical space without genetic alteration.
3. 3. Transient bioelectric modulation of planaria shifts animals to a persistent two-headed regenerative phenotype that survives repeated amputations and exhibits bistability (Durant et al., 2016; Durant et al., 2017; Pezzulo et al., 2021), showing that intercellular signaling — not genomic sequence — encodes target-morphology memory.
4. 4. Newt kidney tubules maintain a constant target diameter whether constructed by the normal ~8 cells, by fewer very large cells, or by a single cell that bends around itself using cytoskeletal mechanisms (Fankhauser, 1945a; 1945b), satisfying William James's 1890 definition of intelligence as achieving the same goal by multiple means.
5. 5. A Hopfield network stores and retrieves multiple associative memories, solves combinatorial optimization problems, and generalizes to novel solutions through fully distributed Hebbian weight updates with no centralized control or global performance feedback — making it the paper's null model for basal collective cognition.
6. 6. Individual reinforcement learning acting on agents with modifiable connections to neighbors produces system-level dynamics equivalent to unsupervised Hebbian learning at the collective scale, meaning collective-level intelligence can emerge without collective-level selection or a unified reward signal (Watson et al., 2011a; Power et al., 2015).
7. 7. The paper hypothesizes that evolutionary transitions in individuality (ETIs) correspond specifically to the emergence of deep interaction structures — multi-layer, non-linearly separable functional relationships — rather than single-layer recurrent networks, and that this 'chunking' of search space is what constitutes a new level of biological individuality (Watson et al., 2022).
8. 8. Cancer is framed as a failure of bioelectric collective intelligence: disconnection from tissue-level bioelectric networks releases cells from higher-order morphogenetic goals, and forcing bioelectrical coupling among cells with strong oncogenic mutations is sufficient to suppress cancer phenotypes and restore large-scale morphogenetic control (Chernet and Levin, 2013).
9. 9. A replication-ready modeling approach used throughout is the 'self-modelling dynamical system' variant of the Hopfield network (rHN-s, Watson et al., 2011a), in which slow Hebbian learning is applied to connection weights while state variables explore locally optimal solutions, producing progressive improvement in solution quality without any external teacher or global error signal.
10. 10. An open question the paper explicitly raises is which specific biological networks — among gene-regulatory, metabolic, bioelectric, and morphogen-diffusion networks — actually meet the non-trivial conditions for distributed connectionist learning, and to what degree this variation explains observed differences in collective intelligence across biological systems.

Peer brief — for seminar discussion

Watson and Levin develop a synthetic theoretical framework — evolutionary connectionism — proposing that the organizational requirements for evolutionary individuality, organismic individuality, and cognition are coextensive rather than categorically distinct phenomena. Working from connectionist models of neural learning (specifically the Hopfield network, single- and multi-layer Perceptrons, deep belief networks, and deep auto-encoders), they argue that the same algorithmic principles governing information integration and associative memory in artificial neural networks apply to gene-regulatory networks, ecological community networks, bioelectric morphogenetic networks, and social networks when connections are suitably arranged. The unifying conjecture is that causal structures producing fitness differences that belong to a collective rather than its parts — the hallmark of an evolutionary unit per Okasha (2006) — are identical to those producing the information integration and coordinated action characteristic of individual cognition. The load-bearing empirical anchors are drawn from developmental biology. Xenopus tadpoles with eyes ectopically relocated to their tails still support light-mediated learning because the eye primordia wire optic nerves to the spinal cord rather than the brain, illustrating goal-directed navigation of anatomical morphospace. Planaria can be stably shifted to a two-headed regenerative phenotype through transient bioelectric manipulation, with the phenotype persisting through subsequent amputations and displaying bistability consistent with a cognitive pattern-memory circuit (Durant et al., 2016; Durant et al., 2017; Pezzulo et al., 2021) — all in cells with unaltered genomes. Newt kidney tubule diameter is maintained across cell-size manipulations ranging from normal ~8-cell assemblies down to a single large cell bending on its cytoskeleton (Fankhauser, 1945a; 1945b), satisfying James's 1890 criterion for intelligence. The paper's core prediction is that evolutionary transitions in individuality correspond to the induction of deep interaction structures — specifically, non-linearly separable functional relationships requiring multi-layer architecture — rather than shallow single-level recurrent networks. This implies that ETIs are mechanistically instances of deep model induction, and that the 'chunking' of search space observed when neural networks acquire deep representations is the same process by which new levels of biological individuality spontaneously emerge. An alternative theoretical vehicle the paper could have used but does not is active inference / free-energy minimization (Friston), which shares the goal of unifying cognition and biological self-organization but grounds individuality in variational inference rather than connectionist learning dynamics. The method introduced — applying distributed Hebbian learning to non-neural biological networks as a formal model of bottom-up collective intelligence emergence — is directly applicable to ecological assembly models (Power et al., 2015) and morphogenetic bioelectric simulations (Manicka and Levin, 2019b). A critical reader would press hardest on the following: the paper treats the formal equivalence between evolutionary dynamics and connectionist learning as licensing the inference that biological systems actually implement cognition, but formal equivalence between algorithmic descriptions does not establish mechanistic identity. Showing that gene-regulatory or bioelectric networks are isomorphic to Hopfield dynamics under certain parameterizations does not demonstrate that those networks are, in any non-metaphorical sense, storing associative memories or generalizing — it may simply redescribe known regulatory robustness in a different vocabulary. The empirical examples (planarian bistability, ectopic eyes) are compelling as demonstrations of morphogenetic goal-directedness, but none of them directly test the connectionist learning hypothesis; they are consistent with purely feed-forward developmental programs as much as with learned attractors. The paper acknowledges this gap but leaves as future work the identification of which biological networks actually meet the non-trivial conditions for distributed learning to operate.

Methods (1)

• Gradient method
  Optimization technique that computes weight changes by following the gradient of an error function; contrasted with evolutionary stochastic search.

Findings (14)

• Transient bioelectrical modulation of body-wide pattern memory circuits in planaria can shift them from one-headed to persistent two-headed state, persisting through amputation rounds until reset with different manipulation.

  Experimental evidence that organism-scale goals can be rewritten through physiological signals without genetic modification; demonstrates bioelectricity as cognitive medium.

• Newt kidney tubule cells produce correct tubule diameter using fewer cells when cell size is enlarged; a single enlarged cell can loop to achieve the same diameter (Fankhauser 1945).

  Shows multi-scale anatomical homeostasis using different cellular mechanisms.

• Ectopic eyes formed in frog embryo tails still function; eye primordia cells succeed in forming an eye, optic nerve, and correct neural connections despite abnormal spatial context.
• Genetically wild-type Girardia dorotocephala flatworms develop different species-specific head anatomies upon gap junctional blockade (Emmons-Bell et al. 2015).

  Demonstrates that gap junctional communication determines species-specific organs without genetic change.

• Disconnection from bioelectric tissue networks enables cancer progression; forced bioelectric coupling suppresses cancer phenotypes despite oncogenic mutations.
• Hopfield network can store multiple patterns, recall them via content-addressable memory, and generalise to novel patterns with the same underlying structure.

  Summary of known Hopfield network capabilities used as a model for collective computation.

• Bioelectric signatures control morphogenetic target patterns; transient bioelectrical modulation in planaria produces persistent two-headed phenotype.
• Developing Xenopus tadpoles can attain normal anatomical outcome despite starting with craniofacial organs scrambled or with wrong number of cells.

  Evidence of morphogenetic problem-solving and anatomical homeostasis across serious perturbations; demonstrates collective intelligence in development.

• Cancer phenotypes can be suppressed by forcing bioelectrical connections among cells, overriding oncogenic mutations (Chernet & Levin 2013).

  Shows that restoring bioelectric cohesion can override single-cell goals.

• Ectopic eyes in the tails of Xenopus tadpoles allow the animals to see and connect optic nerve to spinal cord (Blackiston & Levin 2013).

  Demonstrates competence of eye primordia to achieve function in novel locations.

Claims (46)

• The kind of relationships necessary to produce evolutionary individuality – the generation and heritability of fitness differences at the collective level – are the same as those required to produce organismic individuality – the information integration and collective action characteristic of a self.

  Core conjecture linking evolutionary and organismic individuality.

• The key to being an individual is to have a functional structure in which diverse experiences across its components are bound together in a way that generates causal relationships and composite memories that belong to the higher space of the individual and not its components.

  Defines individuality in terms of information binding across parts.

• Bottom-up learning creates a non-decomposable whole (attractors that are non-linearly separable functions of the inputs and depend on the system’s own internal history), which means that credit assignment or reward at the level of individual parts becomes ineffective.

  Explains how collective cognition becomes irreducible to parts.

• Harnessing the native capability of collective intelligence in the service of biomedicine or bioengineering will require a much better understanding of how to identify, characterise and motivate emergent agents in anatomical, physiological and transcriptional spaces.

  Identifies a key future research direction.

• Organismic individuality evolved through a bottom-up process of collective intelligence, resulting in information integration and coordinated action so well-organised that we observe a new level of organismic and evolutionary individuality.

  Historical narrative for the origin of individuals.

• If intercellular signalling (not genes) is the cognitive medium of a morphogenetic individual, it should be possible to exploit tools of neuroscience to read, interpret and re-write its information content for predictive control.

  Strong prediction from basal cognition framework; validated in planaria and other species through bioelectric manipulation.

• Target morphology shifts occur despite the fact that all of the individual cells have unaltered normal genomes, showing that competent subunits can be pushed to implement diverse organism-scale goals by physiological signals.

  Highlights the non-genetic control of large-scale anatomy.

• To be a bona fide evolutionary unit, a collective must exhibit heritable variation in reproductive success that belongs properly to the collective level – over and above the sum of that exhibited by its component parts.

  Formal requirement for evolutionary individuality.

• Hebbian learning in a self-modelling dynamical system effects a positive feedback on correlations; the more things co-occur, the more the connection between them changes to make them more likely to co-occur in future.

  Describes the self-reinforcing nature of Hebbian learning in networks.

• Understanding what kind of relationships instantiate biological individuality is of great importance to synthetic bioengineering, regenerative medicine, exobiology, robotics and artificial intelligence.

  Argues for practical applications of the theoretical framework.

Hypotheses (7)

• Connectionist learning principles apply equally to evolutionary systems where variation and selection alter network organization, instantiating distributed learning at evolutionary timescales.
• We hypothesise that ecological models fall short of demonstrating spontaneous evolution of a new level of individuality because they are single-level networks of symmetric interactions.

  Explains limitation of current ecological connectionist models.

• Functional relationships necessary for evolutionary individuality are the same as those required for organismic individuality and constitute cognitive architectures.
• Learning neural networks can enable 'chunking' and rescale problem-solving to higher organizational levels, a mechanism intrinsic to transitions in individuality.
• Evolutionary transitions in individuality correspond to deep interaction structures or perhaps other mechanisms of multi-scale dynamics.

  Concrete proposal about necessary architecture for ETIs.

• We hypothesize that this rescaling of the problem-solving search process is intrinsic to transitions in individuality.

  Hypothesis about chunking and ETIs.

• We hypothesize that evolutionary individuality, organismic individuality and cognition are coextensive.

  Strong unification proposal.

Questions (6)

• What kinds of functional relationships and specific organisation turn a collective that is not intelligent into a collective that is?

  Central research question organizing the paper; addresses necessary and sufficient conditions for collective intelligence.

• What kinds of interaction structures are necessary for what kind of collective intelligence and how can these structures emerge?

  Key unresolved question; identifies critical knowledge gaps for understanding collective intelligence across scales.

• What kinds of functional relationships and interaction structures are necessary to turn a collective into an intelligent system?
• How does scaling of reward dynamics bind subunits into intelligent collectives that better navigate novel problem spaces?

  Question linking reward scaling to collective problem-solving improvement.

• How do the parts discern which of their actions should be reinforced?

  Core credit assignment question for distributed systems.

• What kind of cognition can such networks exhibit?

  Question about the cognitive capabilities of distributed networks.

Related work— refs + corpus + external arXiv

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

• Design for an Individual: Connectionist Approaches to the Evolutionary Transitions in Individuality

  in corpus

  2022
  ≈ 88%
• Collective intelligence: A unifying concept for integrating biology across scales and substrates

  in corpus

  2024
  ≈ 87%
• Darwin's agential materials: evolutionary implications of multiscale competency in developmental biology

  in corpus

  2023
  ≈ 87%
• AI: a Bridge toward Diverse Intelligence and Humanity’s Future

  in corpus

  2024
  ≈ 87%
• Developmental Bioelectricity: the cognitive glue enabling evolutionary scaling from physiology to mind

  in corpus

  2023
  ≈ 87%
• Evolutionary connectionism: algorithmic principles underlying the evolution of biological organisation in evo-devo, evo-eco and evolutionary transitions

  cited

  2016
  ≈ 86%
• The biogenic approach to cognition

  in corpus

  2005
  ≈ 86%
• The computational boundary of a 'self': developmental bioelectricity drives multicellularity and scale-free cognition

  in corpus

  2019
  ≈ 86%
• Generalizing frameworks for sentience beyond natural species

  in corpus
  ≈ 84%
• Artificial Theory of Mind and Self-Guided Social Organisation

  Jaime Ruiz-Serra, Catherine Drysdale Michael S. Harr\'e

  2024
  ≈ 84%
• Endless forms most beautiful 2.0: teleonomy and the bioengineering of chimaeric and synthetic organisms

  in corpus

  2023
  ≈ 84%
• Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds

  in corpus

  2022
  ≈ 84%
• Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity

  Michael Levin Giovanni Pezzulo

  2026
  ≈ 83%
• The Machine Consciousness Hypothesis

  in corpus
  ≈ 83%
• The Physical Basis of Prediction: World Model Formation in Neural Organoids via an LLM-Generated Curriculum

  Brennen Hill

  2025
  ≈ 83%
• The scaling of goals via homeostasis: an evolutionary simulation, experiment and analysis

  Johanna Bischof, Jennifer V. LaPalme, and Michael Levin Leo Pio-Lopez

  2022
  ≈ 83%
• Collective Intelligence for Deep Learning: A Survey of Recent Developments

  Yujin Tang David Ha

  2022
  ≈ 83%
• Multiple ways to implement and infer sentience

  in corpus
  ≈ 83%
• Cognition coming about: self-organisation and free-energy

  Maxwell Ramstead, Axel Constant, Karl Friston Ines Hipolito

  2020
  ≈ 83%
• Role of Morphogenetic Competency on Evolution

  Lakshwin Shreesha

  2023
  ≈ 82%
• Intelligence Sequencing and the Path-Dependence of Intelligence Evolution: AGI-First vs. DCI-First as Irreversible Attractors

  Andy E. Williams

  2025
  ≈ 82%
• The evolution of phenotypic correlations and "developmental memory"

  cited

  2014
  ≈ 82%
• Dissociated Neuronal Cultures as Model Systems for Self-Organized Prediction

  Zhuo Zhang, Dai Akita, Tomoyo Isoguchi Shiramatsu, Zenas Chao, Hirokazu Takahashi Amit Yaron

  2025
  ≈ 82%
• Self-Improvising Memory: A Perspective on Memories as Agential, Dynamically Reinterpreting Cognitive Glue

  in corpus

  2024
  ≈ 82%
• Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems

  Xinfeng Li, Jiayi Zhang, Jinlin Wang, Tanjin He, Sirui Hong, Hongzhang Liu, Shaokun Zhang, Kaitao Song, Kunlun Zhu, Yuheng Cheng, Suyuchen Wang, Xiaoqiang Wang, Yuyu Luo, Haibo Jin, Peiyan Zhang, Ollie Liu, Jiaqi Chen, Huan Zhang, Zhaoyang Yu, Haochen Shi, Boyan Li, Dekun Wu, Fengwei Teng, Xiaojun Jia, Jiawei Xu, Jinyu Xiang, Yizhang Lin, Tianming Liu, Tongliang Liu, Yu Su, Huan Sun, Glen Berseth, Jianyun Nie, Ian Foster, Logan Ward, Qingyun Wu, Yu Gu, Mingchen Zhuge, Xinbing Liang, Xiangru Tang, Haohan Wang, Jiaxuan You, Chi Wang, Jian Pei, Qiang Yang, Xiaoliang Qi, Chenglin Wu Bang Liu

  2025
  ≈ 82%
• Global adaptation in networks of selfish components: emergent associative memory at the system scale

  cited

  2011
  ≈ 82%
• Isomorphic Functionalities between Ant Colony and Ensemble Learning: Part III -- Gradient Descent, Neural Plasticity, and the Emergence of Deep Intelligence

  Gregory Babbitt, Yuval Levental Ernest Fokou\'e

  2026
  ≈ 82%
• LLM-Assisted Iterative Evolution with Swarm Intelligence Toward SuperBrain

  Pedro Carvalho Brom, Lucas Ramson Siefert Li Weigang

  2025
  ≈ 82%
• Rethinking Cognition: Morphological Info-Computation and the Embodied Paradigm in Life and Artificial Intelligence

  Gordana Dodig-Crnkovic

  2024
  ≈ 82%
• On the link between conscious function and general intelligence in humans and machines

  Kai Arulkumaran, Shuntaro Sasai, Ryota Kanai Arthur Juliani

  2022
  ≈ 82%

+20 more

Similar preprints — Semantic Scholar

Cited by (1)

• Cognitive glues are shared models of relative scarcities: the economics of collective intelligence

  Levin and Lyons argue that the price system in market economies is not merely a resource-allocation device but the canonical instantiation of what they term a *cognitive glue* — a coordinating afforda