BioelectricaAlterspan Labs
Self-organisation  ·  Article

Exploratory adaptation and random-network search

Exploratory adaptation and random-network search covers work that uses randomly built networks, either as a starting point for search or as a null baseline. Schreier, Soen and Brenner (2017) simulated large random networks and suggested that successful adaptation owes much to how common fixed-point dynamics are in such networks.4 Other authors ask what random networks can store or learn, and compare them with evolved or biological ones.17

Earliest held
2012, McGregor S, Vasas V, Husbands…
Most discussed in
Exploratory adaptation in large random…, 2017
In the library
36 passages in 6 works
Rewritten
2026-10-04
18501875190019251950197520002025McGregor S, Vasas V, Husbands P…, 20121McGregor S, Vasas V, Husbands P…, 20122Lobo D, Levin M, 20153Schreier HI, Soen Y, Brenner N, 20174Schreier HI, Soen Y, Brenner N, 20175Manicka S, Levin M, 20196Biswas S, Clawson W, Levin M, 20227Biswas S, Clawson W, Levin M, 20228Bongard J, Levin M, 20239
01

Relaxation to fixed points

Schreier, Soen and Brenner (2017) simulated many independent networks in each topological ensemble. For each ensemble they computed the fraction of networks that relaxed within a given time window, starting from random initial conditions.5 The measure had no constraint, feedback or random walk in connection strengths. They ran networks up to N=10,000 and found topology-dependent differences that matched the ability for exploratory adaptation seen earlier in the paper. They concluded that a high abundance of networks with fixed points makes a substantial contribution to successful adaptation.4

02

Random starts for search

Other authors use random networks as the first step of a search. Lobo and Levin (2015) searched for regulatory networks with an evolutionary algorithm. The initial population held random networks with random parameters and regulations, and evolution continued until a network with zero error was found.3 Manicka and Levin (2019) trained bioelectric networks by backpropagation, which starts from a random network and gradually tweaks its parameters. They did not claim that organisms learn this way.6

03

What random networks store

McGregor, Vasas, Husbands and Fernando (2012) compared random chemical networks with evolved ones. They noted that randomly constituted systems can often accumulate information in a usable fashion, as in reservoir machines.1 Random networks stored information about the rate of control pulses, though less than evolved networks did. They lacked the machinery to turn that information into an appropriate response.1 A regression model fitted the stored information well for an easy task. For the subtler AB-BA task the fit for random networks was very low, 0.06.2

04

Random networks as baseline

Biswas, Clawson and Levin (2022) tested learning-like memory in biological and random transcriptional networks. Most memory types were more common in biological networks, and no-memory cases were far more common in random ones, which gave a null baseline.7 Both kinds of network could add a second UCS memory, serially or in parallel.8 Bongard and Levin (2023) summarised this, reporting many learning examples in biological networks and many fewer in random controls, and suggested evolution enriches for the property.9

SourcesEach quotation was checked word for word against the passage it opens.
  1. randomly constituted systems can often accumulate information in a usable fashionMcGregor S, Vasas V, Husbands P, Fernando C, 2012 · Evolution of associative learning in chemical networks · open at passage 33
  2. For the AB-BA task, which requires accumulating more subtle information, the quality of fit of the regression model for random networks is very low (0.06)McGregor S, Vasas V, Husbands P, Fernando C, 2012 · Evolution of associative learning in chemical networks · open at passage 30
  3. The initial population comprises random networks with random parameters and regulations between the phenotypic products, the wound product, and the genetic and pharmacological perturbed productsLobo D, Levin M, 2015 · Inferring regulatory networks from experimental morphological phenotypes: a… · open at passage 51
  4. This suggests that a substantial contribution to successful adaptation is indeed provided by a high abundance of networks exhibiting fixed points in their dynamics.Schreier HI, Soen Y, Brenner N, 2017 · Exploratory adaptation in large random networks · open at passage 27
  5. we computed, for each topological ensemble, the fraction of networks supporting relaxation within a given time window, starting with random initial conditions.Schreier HI, Soen Y, Brenner N, 2017 · Exploratory adaptation in large random networks · open at passage 27
  6. BP is a training method, often used in machine learning, that starts with a random network (defined with a random set of parameters) and gradually tweaks the parameters using gradient descentManicka S, Levin M, 2019 · Modeling somatic computation with non-neural bioelectric networks · open at passage 17
  7. Interestingly, the average number of no memory cases (no memory or trivial S-R relationships) is much larger in random networks, providing an important null baseline.Biswas S, Clawson W, Levin M, 2022 · Learning in Transcriptional Network Models: Computational Discovery of… · open at passage 29
  8. both biological and random networks were able to add an additional UCS to a UCS memory in both cases of serial or parallel memoriesBiswas S, Clawson W, Levin M, 2022 · Learning in Transcriptional Network Models: Computational Discovery of… · open at passage 23
  9. We found numerous examples of learning capacity in biological networks and many fewer in control random networks, suggesting that evolution is enriching for this propertyBongard J, Levin M, 2023 · There's Plenty of Room Right Here: Biological Systems as Evolved, Overloaded… · open at passage 46
Linked ideas
Exploratory adaptation is a search by random variation and retention, much as trial and error is, though here the search runs in connection strengths.
Evolvabilityrelated to
Random networks that often relax to fixed points bear on how readily evolved systems find workable solutions.
Basal cognitionrelated to
Learning found in gene regulatory and chemical networks is offered as evidence for cognition-like function below the level of nervous systems.
Conditioned reflexprecursor of / follows
Biswas and colleagues test networks for Pavlovian UCS-CS-response memory, the classical associative learning scheme.
Relaxation of a network to a fixed point is a simple form of settling to a stable state.
Where it is discussedPassages matching exploratory adaptation, random network, gene regulatory network adaptation, relaxation to fixed point
2017Exploratory adaptation in large random networks · Schreier HI, Soen Y, Brenner N16
2022Learning in Transcriptional Network Models: Computational Discovery of… · Biswas S, Clawson W, Levin M12
2012Evolution of associative learning in chemical networks · McGregor S, Vasas V, Husbands P…5
2019Modeling somatic computation with non-neural bioelectric networks · Manicka S, Levin M1
2015Inferring regulatory networks from experimental morphological phenotypes: a… · Lobo D, Levin M1
2023There's Plenty of Room Right Here: Biological Systems as Evolved, Overloaded… · Bongard J, Levin M1