Learning Time-Varying Connectomes for Stable Behavior Across Development
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Learning Time-Varying Connectomes for Stable Behavior Across Development".
Marcus: Here's the gist of the paper. Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry, and intrinsic dynamics change across development The results demonstrate that a shared spatiotemporal time-varying connectivity program can produce recurrent neural networks that preserve behavior across substantial developmental changes This provides a framework for studying how functional stability can emerge without requiring structural or dynamical stability Developmental Modeling and Hypothesis The study investigated whether a single compact generative encoding could produce functional recurrent controllers across developmental changes The developmental substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants The authors hypothesized that developing connectomes possess compressible spatiotemporal structure that can be represented directly as a function of neural geometry and developmental time They tested this using two geometrically distinct developmental substrates and two visuomotor tasks Optimization Strategy The researchers formulated developmental functional stability as a trajectory-level optimization problem They evolved a compact connectivity generator that receives neural geometry and developmental time as inputs to generate the connectivity of a recurrent controller for the substrate at any point of development Optimization was performed across multiple developmental snapshots, and performance was evaluated across the full developmental interval rather than just at maturity Performance Across Development Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint In contrast, controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval This stability was observed across both substrate geometries and both visuomotor tasks Connectome Stability Mechanisms The results show that stability did not require an invariant connectome or a single trajectory of sparsity, weight sign, or mean connection strength Some controllers benefited from later-born neurons, whereas others remained effective when those neurons were removed The evolved CPPNs nevertheless almost always incorporated developmental time into their generated connectivity The study shows that a compact connectivity generator conditioned on developmental time can preserve visuomotor function while a recurrent neural substrate undergoes repeated topological changes and continuous changes in geometry and intrinsic dynamics These results extend generative neuroevolution from density scaling and morphology-conditioned control to functional connectivity across changing neural substrates The framework provides a basis for studying artificial neural networks as developing systems and for asking which growth programs, substrate organizations, and biological constraints allow behavior to remain effective during neural development Page 3 --- Learning Time-Varying Connectomes for Stable Behavior Across Development Christof Fehrman1, Jacob J.
Ines: Morra1 and Eva A.
Marcus: Naumann1 1 Neurobiology, Duke University, Durham, 27710, NC, USA.
Ines: Abstract Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry, and intrinsic dynamics change across development To model developing visuomotor circuits and behavior we investigated whether a single compact generative encoding could produce functional recurrent controllers across such changes Building on the HyperNEAT algorithm we represented this program as a compositional pattern-producing network that received developmental time as an input to generate age-specific connectivity The developmental substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants We tested this approach using two geometrically distinct developmental substrates and two visuomotor tasks optomotor responses and phototaxis Controllers optimized at only the initial or mature substrate generally failed when evaluated at other ages In contrast controllers optimized across multiple developmental snapshots maintained substantially higher and more consistent performance across trained and unseen ages Optimization at only three developmental snapshots was often sufficient to produce effective behavior across the tested developmental interval The resulting connectomes changed substantially with age and did not converge on a single trajectory of sparsity or connection-weight statistics Controllers also differed in whether they relied on neurons born later in development These results demonstrate that a shared spatiotemporal time-varying connectivity program can produce recurrent neural networks that preserve behavior across substantial developmental changes Together these results provide a framework for studying how functional stability can emerge without requiring structural or dynamical stability Page 4 --- Introduction Neural development involves the birth and maturation of neurons, the formation and elimination of synapses, changes in neuronal position and morphology, and continuing refinement of intrinsic and synaptic properties
1, 2: Neural development should therefore be understood not merely as the construction of a mature nervous system but as a process that sustains useful computation along the path to maturity Preserving behavioral function across development does not require the underlying circuit to remain unchanged Functional continuity cannot be inferred from architectural similarity or parameter preservation alone and must instead be evaluated at the level of behavior Representing an entire developmental trajectory of connectomes creates an additional challenge This creates a “genomic bottleneck” through which the information required to construct much larger neural systems must pass Hypernetworks provide an analogous computational strategy in which one network generates parameters for another, typically larger network This perspective motivates the more specific hypothesis that developing connectomes possess compressible spatiotemporal structure that can be represented directly as a function of neural geometry and developmental time Because each developmental stage modifies the preceding network both connectome structure and the neural activity it supports are expected to exhibit temporal continuity rather than vary independently across ages Page 5 --- Sensory Obs.
Paper summary: Ines: So we're looking at this paper, "Learning Time-Varying Connectomes for Stable Behavior Across Development." The main idea here is that neural circuits have to keep working effectively while their size, shape, and how fast things happen change as they develop.
Marcus: Right. And this study investigates if you can use a single compact generative encoding to create recurrent controllers that stay functional across all those developmental changes. It's trying to figure out if you need structural or dynamical stability for behavior to remain stable during development.
Yuki: From a population genetics view, it’s interesting because we see how species adapt over time, and this paper suggests the underlying computational structure can be flexible enough to handle those massive internal shifts. It's about how functional continuity emerges without needing a fixed architecture or fixed dynamics across ages.
Ines: Exactly. The researchers hypothesized that developing connectomes have a compressible spatiotemporal structure that they can represent directly as a function of neural geometry and developmental time. They tested this using two different geometric layouts for the brain and two specific tasks, optomotor responses and phototaxis.
Marcus: And the optimization strategy they used was to treat developmental functional stability as a trajectory-level optimization problem. They evolved a compact connectivity generator that takes neural geometry and developmental time as inputs to generate the connectivity for the substrate at any point in development.
Yuki: That sounds like they were trying to build a system that learns how to generate connections across a whole timeline, not just once at the end of growth. It’s about training it across different ages.
Ines: And their performance results were really telling. Controllers optimized only at one developmental endpoint generally failed when you tried to test them at other ages. But controllers optimized across development maintained substantially higher and more consistent fitness over the entire developmental interval, regardless of the substrate geometry or task they were tested on.
Marcus: So they found that stability didn't rely on having an invariant connectome, or just one trajectory of sparsity or weight sign. Some controllers even did better when they relied on neurons born later in development, while others worked fine even if those later neurons were removed.
Yuki: That suggests that the way the network learns to change its structure over time is more important than locking down a specific set of connections at any single moment. It points toward a dynamic generation process being key for survival during development.
Ines: So, in conclusion, the paper "Learning Time-Varying Connectomes for Stable Behavior Across Development" shows that you can use a compact connectivity generator conditioned on developmental time to preserve visuomotor function even when the neural substrate undergoes repeated topological changes and continuous shifts in geometry and dynamics.
Marcus: It’s a framework that extends generative neuroevolution from just density scaling or morphology-conditioned control to functional connectivity across changing neural substrates. This gives us a way to study artificial neural networks as developing systems, asking what growth programs allow behavior to remain effective during development.
Yuki: For the wider world, this suggests that biological systems might be inherently designed with a mechanism for functional stability that isn't dependent on perfect structural rigidity from day one. It opens up questions about which growth programs and substrate organizations allow behavior to persist through neural maturation.
Conclusion: Ines: So we're talking about "Learning Time-Varying Connectomes for Stable Behavior Across Development" by Fehrman, Morra, and Naumann. This paper looks at how a single program can keep a neural circuit working even as its physical structure completely changes during growth.
Marcus: Yeah, what they did is they built this generative encoding that takes developmental time as an input to create the network's connections for any point in time. It’s not about one snapshot; it’s about learning how to generate the whole timeline of connections.
Yuki: From a population genetics standpoint, this suggests that functional continuity isn't tied to a static blueprint. Instead, it seems like there's an underlying temporal rule that guides the structure toward the right behavior across different ages.
Ines: Exactly. It means we don't need to track every single synapse being born or killed perfectly at every moment in development for the circuit to succeed functionally. Stability emerges from this shared spatiotemporal program, not from having a fixed architecture.
Marcus: And they showed that this stability didn't require the network's structure—the actual connections—to be constant, or even for things like sparsity or connection strength to follow a single simple path. Some controllers actually did better when they relied on neurons that were born later in development.
Yuki: That’s interesting because it implies the developmental program is more flexible than we thought, allowing different structural programs to lead to the same functional outcome. It supports the idea that behavior can be preserved through diverse connectivity paths.
Ines: So what this means for us is that when we look at real development, we shouldn't just be looking for one perfect structure, but rather understanding the rules of generation itself—what constraints allow a network to remain effective across massive structural changes.
Marcus: And this paper offers a way to model artificial neural networks as developing systems, showing how they can achieve functional continuity without needing structural stability at every stage. We’ll dig into what these kinds of generative programs mean for designing next-generation AI architectures next.
Christof Fehrman, Jacob J. Morra, Eva A. Naumann
Neurobiology, Duke University
q-bio.NC, cs.NE
Submitted: 2026-11-04
Updated: 2026-11-04
Comments: 29 pages, 6 figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry…
Key concepts
- Developmental Substrate
- This refers to the biological constraints of neural development, such as when neurons are born (birth), where they are placed (position), how they process sensory input, and their inherent speed or timing (intrinsic time constants). The study used different versions of this substrate to simulate different stages of brain growth.
- Time-Varying Connectivity Program
- This is a single generative encoding designed to produce the specific connections (connectome) needed for a neural network at any given point in development. Instead of one fixed structure, this program changes its output based on the developmental time input, allowing it to generate age-specific connectivity patterns.
- Functional Stability
- This is the ability of a neural circuit to continue performing its intended task (like motor control) even when its underlying physical structure—its size, shape, or connections—is continuously changing. The paper shows this can emerge from a time-varying program rather than requiring the physical structure itself to remain static.
Summary
Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry, and intrinsic dynamics change across development The results demonstrate that a shared spatiotemporal time-varying connectivity program can produce recurrent neural networks that preserve behavior across substantial developmental changes This provides a framework for studying how functional stability can emerge without requiring structural or dynamical stability
Developmental Modeling and Hypothesis
The study investigated whether a single compact generative encoding could produce functional recurrent controllers across developmental changes The developmental substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants The authors hypothesized that developing connectomes possess compressible spatiotemporal structure that can be represented directly as a function of neural geometry and developmental time They tested this using two geometrically distinct developmental substrates and two visuomotor tasks
Optimization Strategy
The researchers formulated developmental functional stability as a trajectory-level optimization problem They evolved a compact connectivity generator that receives neural geometry and developmental time as inputs to generate the connectivity of a recurrent controller for the substrate at any point of development Optimization was performed across multiple developmental snapshots, and performance was evaluated across the full developmental interval rather than just at maturity
Performance Across Development
Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint In contrast, controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval This stability was observed across both substrate geometries and both visuomotor tasks
Connectome Stability Mechanisms
The results show that stability did not require an invariant connectome or a single trajectory of sparsity, weight sign, or mean connection strength Some controllers benefited from later-born neurons, whereas others remained effective when those neurons were removed The evolved CPPNs nevertheless almost always incorporated developmental time into their generated connectivity
Conclusion and Implications
The study shows that a compact connectivity generator conditioned on developmental time can preserve visuomotor function while a recurrent neural substrate undergoes repeated topological changes and continuous changes in geometry and intrinsic dynamics These results extend generative neuroevolution from density scaling and morphology-conditioned control to functional connectivity across changing neural substrates The framework provides a basis for studying artificial neural networks as developing systems and for asking which growth programs, substrate organizations, and biological constraints allow behavior to remain effective during neural development
--- Page 3 ---
Learning Time-Varying Connectomes for Stable Behavior Across Development
Christof Fehrman1, Jacob J. Morra1 and Eva A. p. 0 Learning Time-Varying Connectomes for Stable
Behavior Across Development
Christof Fehrman1*, Jacob J. Morra1 and Eva A. Morra1 and Eva A. Naumann1
1*Neurobiology, Duke University, Durham, 27710, NC, USA.Where the paper says it
Where the paper says it
Abstract Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry, and intrinsic dynamics change across development To model developing visuomotor circuits and behavior we investigated whether a single compact generative encoding could produce functional recurrent controllers across such changes Building on the HyperNEAT algorithm we represented this program as a compositional pattern-producing network that received developmental time as an input to generate age-specific connectivity The developmental substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants We tested this approach using two geometrically distinct developmental substrates and two visuomotor tasks optomotor responses and phototaxis Controllers optimized at only the initial or mature substrate generally failed when evaluated at other ages In contrast controllers optimized across multiple developmental snapshots maintained substantially higher and more consistent performance across trained and unseen ages Optimization at only three developmental snapshots was often sufficient to produce effective behavior across the tested developmental interval The resulting connectomes changed substantially with age and did not converge on a single trajectory of sparsity or connection-weight statistics Controllers also differed in whether they relied on neurons born later in development These results demonstrate that a shared spatiotemporal time-varying connectivity program can produce recurrent neural networks that preserve behavior across substantial developmental changes Together these results provide a framework for studying how functional stability can emerge without requiring structural or dynamical stability
--- Page 4 ---
Introduction Neural development involves the birth and maturation of neurons, the formation and elimination of synapses, changes in neuronal position and morphology, and continuing refinement of intrinsic and synaptic properties [1, 2] Neural development should therefore be understood not merely as the construction of a mature nervous system but as a process that sustains useful computation along the path to maturity Preserving behavioral function across development does not require the underlying circuit to remain unchanged Functional continuity cannot be inferred from architectural similarity or parameter preservation alone and must instead be evaluated at the level of behavior Representing an entire developmental trajectory of connectomes creates an additional challenge This creates a “genomic bottleneck” through which the information required to construct much larger neural systems must pass Hypernetworks provide an analogous computational strategy in which one network generates parameters for another, typically larger network This perspective motivates the more specific hypothesis that developing connectomes possess compressible spatiotemporal structure that can be represented directly as a function of neural geometry and developmental time Because each developmental stage modifies the preceding network both connectome structure and the neural activity it supports are expected to exhibit temporal continuity rather than vary independently across ages
--- Page 5 ---
Sensory Obs. p. 4 Sensory Obs. Sensory Obs. Sensory Obs. Sensory Obs. Sensory Obs. Action
Action
Action
Developmental Time
T1 T2 T3
X
Y
src
xy
tgtxy
T
wsrc tgt
expr.Where the paper says it
Where the paper says it
Where the paper says it
Fig. 1 Parameterizing the connectome of a developing visuomotor system (A) Three developmental snapshots illustrate the challenge of maintaining function as the neural substrate changes The system must use sensory observations to produce survival-relevant actions while accommodating changes in network topology (e. p. 4 (A) Three developmental
snapshots illustrate the challenge of maintaining function as the neural substrate changes. The system must use sensory observations to produce survival-relevant actions while accommodating changes
in network topology (e.Where the paper says it
--- Page 7 ---
Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint One HyperNEAT evolutionary run showed partial stability across OMR development but this behavior was not representative of the condition and remained less consistent than that of the developmentally aware conditions Controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval Increasing the number of training snapshots generally produced more consistent developmental trajectories, but notably, even Dev. Aware–3pts controllers could interpolate across the intervals separating their training snapshots
--- Page 8 ---
Fish Ring 0 100 200 Generations Best Genome Fitness 0.5 1.0 Mean Fitness
Fig. 2 Results of Task Optimization and Fitness Stability (A) Fitness trajectories across training conditions and substrate–task pairs Dev Aware–3pts received developmental time as an additional CPPN input and were optimized for mean fitness across three, five, or fifteen evenly spaced developmental snapshots respectively Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint In contrast controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval
--- Page 9 ---
Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint In contrast controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval
--- Page 10 ---
Controllers optimized at a single developmental endpoint generally failed to transfer away from that endpoint In contrast controllers optimized across development maintained substantially higher and more consistent fitness over the developmental interval
--- Page 13 ---
Table 1 HyperNEAT substrate parameter values for the Fish and Ring substrates Parameter Fish Ring Developmental interpolation, g(T) T T 2 Lateral assignment of recurrent cohort ranks Outside to inside Inside to outside Middle-cohort birth-window bounds [0.00, 0.60] [0.35, 1.14] Late-cohort birth-window bounds [0. p. 0 Building on the HyperNEAT algorithm, we represented this
program as a compositional pattern-producing network that received developmental time as an input to generate age-specific connectivity. The developmental
substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants. In contrast, HyperNEAT was optimized on the smaller initial
substrate, and the developmentally aware conditions had to maximize mean performance across multiple substrates rather than at a single developmental endpoint. Despite this more demanding objective, the final training fitnesses of the developmentally aware conditions were generally comparable to those of HyperNEAT, with their
relative ordering varying across substrate–task combinations (Fig. In contrast, HyperNEAT was optimized on the smaller initial
substrate, and the developmentally aware conditions had to maximize mean performance across multiple substrates rather than at a single developmental endpoint. Despite this more demanding objective, the final training fitnesses of the developmentally aware conditions were generally comparable to those of HyperNEAT, with their
relative ordering varying across substrate–task combinations (Fig.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 14 ---
Table 2 Allowed source→target projections for the Fish and Ring substrates Substrate Allowed source → target projections Fish Sensory L→ Area 1 R; sensory R→ Area 1 L; Area 2→ Area 3; Area 3→ motor. Ring Sensory L→ Area 2 R; sensory R→ Area 2 L; Area 2R→ motor L,Area1L, and A3L;Area2L→ motorR,Area1R, and A3R. p. 3 Their
CPPN generated morphology-dependent connection weights, neuronal biases, and time
constants. However, the controller substrate itself remained fixed across robot morphologies: the same neurons, neural coordinates, sensory and motor organization, and
set of possible connections were retained.Where the paper says it
--- Page 15 ---
Table 6 Substrate Growth Models (A) Three developmental snapshots (T) of the Fish and Ring substrates Each differs in neuronal migration, neurogenesis schedules, and projection constraints
--- Page 17 ---
Table 3 CPPN encoding and neuroevolution parameters Table 3 CPPN encoding and neuroevolution parameters Each differs in neuronal migration, neurogenesis schedules, and projection constraints
--- Page 18 ---
Table 3 CPPN encoding and neuroevolution parameters Table 3 CPPN encoding and neuroevolution parameters Each differs in neuronal migration, neurogenesis schedules, and projection constraints
--- Page 20 ---
Table 3 CPPN encoding and neuroevolution parameters Table 3 CPPN encoding and neuroevolution parameters Each differs in neuronal migration, neurogenesis schedules, and projection constraints
--- Page 24 ---
Discussion Neural development requires a controller to remain functional while the network performing the computation changes Most developmental encodings evaluate the controller produced after growth This study is among the first to require an artificial recurrent controller to preserve a selected behavior across a developmental trajectory that combines these forms of neural change Performance at untrained ages suggests that the CPPNs captured regularities shared across developmental stages rather than requiring an independently optimized connectome at every age The results also show that stability did not require an invariant connectome or a single trajectory of sparsity, weight sign, or mean connection strength Some controllers benefited from later-born neurons, whereas others remained effective when those neurons were removed Stable behavior could therefore arise through different developmental connectivity programs
--- Page 27 ---
Acknowledgements Research reported in this publication was supported by the National Institutes of Health under Award Number R01 EY033845–01 The content is solely the authors’ responsibility and does not necessarily represent the official views of the National Institutes of Health J.J.M. was supported by a Ruth K. Broad Foundation Postdoctoral Fellowship E.A.N. was supported by the Whitehall and Alfred P. Sloan Foundation Data and Code Availability All data and code used for this project will be publicly released upon publication
--- Page 28 ---
References [1] Goodhill, G.J.: Theoretical models of neural development. p. 1 Neural development has received substantial computational attention, but existing
approaches typically capture only part of this problem. Many models describe developmental changes in synaptic weights, intrinsic excitability, or neural activity while
retaining a fixed set of neurons and allowable connections [7]. Many models describe developmental changes in synaptic weights, intrinsic excitability, or neural activity while
retaining a fixed set of neurons and allowable connections [7]. Such models can reveal
how basic circuit computation changes during development, but cannot directly represent neurogenesis, changing neural geometry, or changes in the fundamental network
topology. Morra1 and Eva A. Naumann1
1*Neurobiology, Duke University, Durham, 27710, NC, USA. 1 Introduction
Neural development involves the birth and maturation of neurons, the formation and
elimination of synapses, changes in neuronal position and morphology, and continuing refinement of intrinsic and synaptic properties [1, 2]. Since an animal cannot wait
until it reaches maturity before interacting successfully with its environment, the nervous system must produce behavior that supports survival throughout these changes. Since an animal cannot wait
until it reaches maturity before interacting successfully with its environment, the nervous system must produce behavior that supports survival throughout these changes. Neural development should therefore be understood not merely as the construction of
a mature nervous system, but as a process that sustains useful computation along the
path to maturity. 06895v1 [q-bio. NC] 4 Nov 2026 For example, phase relationships in the lobster pyloric
motor rhythm remain similar between juvenile and adult animals despite pronounced
changes in neural and body size [3], and neurons established during embryogenesis can
retain important locomotor functions across life stages despite extensive structural
remodeling [4]. Similar network outputs can arise from substantially different combinations of intrinsic and synaptic parameters [5], and developing circuits can recruit
different combinations of homeostatic mechanisms across cell types, perturbations, and
developmental stages [6]. For example, phase relationships in the lobster pyloric
motor rhythm remain similar between juvenile and adult animals despite pronounced
changes in neural and body size [3], and neurons established during embryogenesis can
retain important locomotor functions across life stages despite extensive structural
remodeling [4]. Similar network outputs can arise from substantially different combinations of intrinsic and synaptic parameters [5], and developing circuits can recruit
different combinations of homeostatic mechanisms across cell types, perturbations, and
developmental stages [6]. For example, phase relationships in the lobster pyloric
motor rhythm remain similar between juvenile and adult animals despite pronounced
changes in neural and body size [3], and neurons established during embryogenesis can
retain important locomotor functions across life stages despite extensive structural
remodeling [4]. Similar network outputs can arise from substantially different combinations of intrinsic and synaptic parameters [5], and developing circuits can recruit
different combinations of homeostatic mechanisms across cell types, perturbations, and
developmental stages [6]. For example, phase relationships in the lobster pyloric
motor rhythm remain similar between juvenile and adult animals despite pronounced
changes in neural and body size [3], and neurons established during embryogenesis can
retain important locomotor functions across life stages despite extensive structural
remodeling [4]. Similar network outputs can arise from substantially different combinations of intrinsic and synaptic parameters [5], and developing circuits can recruit
different combinations of homeostatic mechanisms across cell types, perturbations, and
developmental stages [6]. naumann@duke. edu;
Abstract
Neural circuits must maintain functionality to drive effective behavior while their
size, topology, geometry, and intrinsic dynamics change across development. Such models can reveal
how basic circuit computation changes during development, but cannot directly represent neurogenesis, changing neural geometry, or changes in the fundamental network
topology. Other methods explicitly enlarge artificial neural networks, often by adding
neurons or layers through a small number of discrete transformations. Such models can reveal
how basic circuit computation changes during development, but cannot directly represent neurogenesis, changing neural geometry, or changes in the fundamental network
topology. Other methods explicitly enlarge artificial neural networks, often by adding
neurons or layers through a small number of discrete transformations. Learning Time-Varying Connectomes for Stable
Behavior Across Development
Christof Fehrman1*, Jacob J. Morra1 and Eva A. Learning Time-Varying Connectomes for Stable
Behavior Across Development
Christof Fehrman1*, Jacob J. Morra1 and Eva A. Together, these results provide a framework for studying how functional stability
can emerge without requiring structural or dynamical stability. Keywords: Developmental neuroevolution, Embodied agents, Evolutionary
algorithms, Recurrent neural networks
1
arXiv:2611. Together, these results provide a framework for studying how functional stability
can emerge without requiring structural or dynamical stability. Keywords: Developmental neuroevolution, Embodied agents, Evolutionary
algorithms, Recurrent neural networks
1
arXiv:2611. Related work extends structural
plasticity into the agent’s lifetime but focuses on experience- and reward-dependent
adaptation rather than preservation of a selected function across a prescribed developmental trajectory [13]. Together, these approaches leave open whether a controller
can sustain high behavioral performance as its neural substrate changes over developmental time. Related work extends structural
plasticity into the agent’s lifetime but focuses on experience- and reward-dependent
adaptation rather than preservation of a selected function across a prescribed developmental trajectory [13]. Together, these approaches leave open whether a controller
can sustain high behavioral performance as its neural substrate changes over developmental time. Together, these results provide a framework for studying how functional stability
can emerge without requiring structural or dynamical stability. Keywords: Developmental neuroevolution, Embodied agents, Evolutionary
algorithms, Recurrent neural networks
1
arXiv:2611. Together, these results provide a framework for studying how functional stability
can emerge without requiring structural or dynamical stability. Keywords: Developmental neuroevolution, Embodied agents, Evolutionary
algorithms, Recurrent neural networks
1
arXiv:2611. Other methods explicitly enlarge artificial neural networks, often by adding
neurons or layers through a small number of discrete transformations. These transformations may preserve a feedforward network’s function [8], increase its capacity
during training [9], or allow it to acquire new information without erasing existing
representations [10]. Developmental encodings provide a more flexible alternative by specifying local or
generative rules through which network structure emerges. Recent neural developmental programs can add neurons, differentiate cell states, and construct networks with
changing topologies [11, 12]. Such models can reveal
how basic circuit computation changes during development, but cannot directly represent neurogenesis, changing neural geometry, or changes in the fundamental network
topology. Other methods explicitly enlarge artificial neural networks, often by adding
neurons or layers through a small number of discrete transformations. In a layered feedforward network, additional units
can sometimes be introduced through algebraic transformations that preserve a static
input–output mapping [8]. In recurrent neural networks, changes in recurrent connectivity and neuronal time constants can alter transient response times, equilibrium
stability, oscillatory regimes, and attractor dynamics [14, 15]. Learning Time-Varying Connectomes for Stable
Behavior Across Development
Christof Fehrman1*, Jacob J. Morra1 and Eva A. In a layered feedforward network, additional units
can sometimes be introduced through algebraic transformations that preserve a static
input–output mapping [8]. In recurrent neural networks, changes in recurrent connectivity and neuronal time constants can alter transient response times, equilibrium
stability, oscillatory regimes, and attractor dynamics [14, 15]. Representing an entire developmental trajectory of connectomes creates an additional challenge. One approach would be to specify every synapse independently at
every developmental age, but such a representation would be inefficient and a poor
abstraction of biological development. Such models can reveal
how basic circuit computation changes during development, but cannot directly represent neurogenesis, changing neural geometry, or changes in the fundamental network
topology. Other methods explicitly enlarge artificial neural networks, often by adding
neurons or layers through a small number of discrete transformations.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 26 ---
[8] Chen, T., Goodfellow, I., Shlens, J.: Net2net: Accelerating learning via knowledge transfer. arXiv preprint arXiv:1511.05641 (2015) [9] Fehring, L., Lindauer, M., Eimer, T.: Growing with experience: Growing neural networks in deep reinforcement learning. arXiv preprint arXiv:2506.11706 [23] Stanley, K. p. 3 Morphology-conditioned HyperNEAT provides a related precedent. Risi and Stanley [25] conditioned a CPPN on quadruped leg length and evolved controllers that
generalized to intermediate leg lengths not encountered during optimization.Where the paper says it
--- Page 29 ---
[19] Ha, D., Dai, A., Le, Q.V.: HyperNetworks (2016). p. 2 This
creates a “genomic bottleneck” through which the information required to construct
much larger neural systems must pass [18]. Hypernetworks provide an analogous computational strategy in which one network generates parameters for another, typically
larger network [19, 20]. Together, these results provide a framework for studying how functional stability
can emerge without requiring structural or dynamical stability. Keywords: Developmental neuroevolution, Embodied agents, Evolutionary
algorithms, Recurrent neural networks
1
arXiv:2611. Morphology-conditioned HyperNEAT provides a related precedent. Risi and Stanley [25] conditioned a CPPN on quadruped leg length and evolved controllers that
generalized to intermediate leg lengths not encountered during optimization.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 30 ---
[24] Stanley, K.O.: Compositional pattern producing networks: A novel abstraction of development. p. 2 HyperNEAT evolves
compositional pattern-producing networks (CPPNs) that generate neural connectivity
as a function of substrate geometry, allowing a compact encoding to parameterize networks containing many connections. CPPNs were originally proposed as an abstraction
of biological development, in which a compact genotype generates spatial structure
without explicitly simulating cell-by-cell growth [24].Where the paper says it
--- Page 31 ---
[5] Prinz, A.A., Bucher, D., Marder, E.: Similar network activity from disparate circuit parameters. Nature neuroscience 7(12), 1345–1352 (2004)
--- Page 32 ---
[6] Tien, N.-W., Kerschensteiner, D.: Homeostatic plasticity in neural development. Neural development 13(1), 9 (2018)
--- Page 33 ---
[7] Zavitz, D., Ching, S., Goodhill, G.: Modeling normal and abnormal circuit development with recurrent neural networks. Cold Spring Harbor perspectives in biology 17(6), 041507 (2025)
--- Page 34 ---
[8] Chen, T., Goodfellow, I., Shlens, …
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Developmentally aware controllers can maintain substantially higher and more consistent fitness over a developmental interval compared to controllers optimized at a single endpoint, suggesting that optimization across multiple developmental snapshots is sufficient to produce effective behavior across substantial changes.
-
The system can produce qualitatively coherent visuomotor behavior at unseen ages, such as producing oscillatory distance trajectories for phototaxis on the Ring substrate by utilizing an active scanning strategy.
-
The AI system can learn to exploit neurogenesis or rely on neurons present at the beginning of development by calculating Delta F(T), which indicates whether later-born neurons improved or diminished performance compared to a no neurogenesis condition.
-
The generative encoding allows for
functional stability without requiring structural or dynamical stability,
meaning the system can maintain function across substantial changes in connectivity, geometry, and intrinsic dynamics. -
The AI system can be tested on new substrate geometries and tasks by using a shared generative program that incorporates developmental time as an input to generate connectivity rules appropriate for any point of development.
Abstract
Neural circuits must maintain functionality to drive effective behavior while their size, topology, geometry, and intrinsic dynamics change across development. To model developing visuomotor circuits and behavior, we investigated whether a single compact generative encoding could produce functional recurrent controllers across such changes. Building on the HyperNEAT algorithm, we represented this program as a compositional pattern-producing network that received developmental time as an input to generate age-specific connectivity. The developmental substrate prescribed neuronal birth, position, sensory organization, and intrinsic time constants. We tested this approach using two geometrically distinct developmental substrates and two visuomotor tasks: optomotor responses and phototaxis. Controllers optimized at only the initial or mature substrate generally failed when evaluated at other ages. In contrast, controllers optimized across multiple developmental snapshots maintained substantially higher and more consistent performance across trained and unseen ages. Optimization at only three developmental snapshots was often sufficient to produce effective behavior across the tested developmental interval. The resulting connectomes changed substantially with age and did not converge on a single trajectory of sparsity or connection-weight statistics. Controllers also differed in whether they relied on neurons born later in development. These results demonstrate that a shared spatiotemporal time-varying connectivity program can produce recurrent neural networks that preserve behavior across substantial developmental changes. Together, these results provide a framework for studying how functional stability can emerge without requiring structural or dynamical stability.
Sources
- Net2Net: Accelerating Learning via Knowledge Transfer
- Growing with Experience: Growing Neural Networks in Deep Reinforcement Learning
- Evolution Strategies at the Hyperscale
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks