Similar activity from different circuit parameters

Citation: Prinz, A. A., Bucher, D. and Marder, E. “Similar network activity from disparate circuit parameters.” Nature Neuroscience 7, 1345–1352. Published online 21 November 2004; December 2004 issue. DOI: 10.1038/nn1352. Peer-reviewed primary computational study. Read 29 September 2026.

Reading scope and version

The complete eight-page publisher-formatted article was read from an institutional PDF copy hosted at NJIT, archived as papers/consciousness_connectomics/prinz2004_degeneracy.institution.pdf. All 540 lines of the layout-preserving text extraction were read, including the abstract, introduction, all Results/Discussion, complete Methods, equations, both tables, acknowledgments, declarations and all 50 bibliographic entries. All eight pages were also visually inspected, including every panel of Figures 1–6 and all rows of Tables 1–2. Referenced papers and related search results were navigation/bibliographic material, not independently read evidence.

The official Nature page was archived to verify identity, publication metadata and version history. It supplies a subscription preview rather than complete article text; requesting the publisher PDF returned that HTML rather than a PDF. Nature records a 15 May 2006 correction restoring Figure 5 in the HTML presentation. Figure 5 is present in the institutional PDF and was inspected. The PDF metadata gives a modification date of 9 May 2006; this does not establish that the institutional file is byte-identical to a current publisher download.

No separate written scientific supplement is listed or linked on the archived publisher page, and the complete article contains no supplementary-material reference. There is therefore no identified unread supplement for this article. The older single-neuron database paper, detailed inherited channel equations, source code, full network database and original 99-animal recording set were not independently acquired or audited. No simulations were rerun. Exact source URLs, byte counts, SHA-256 hashes and reading scope are in papers/consciousness_connectomics/prinz2004_degeneracy.provenance.json.

Question and central result

Must a small neural circuit have narrowly specified cellular and synaptic parameters to generate a tightly specified motor rhythm? The authors enumerate 20,250,000 versions of a simplified crustacean pyloric circuit, then retain networks producing specified rhythmic features. Many very different parameter combinations pass those output tests.

The result is a concrete demonstration of non-unique parameter inference from a bounded set of output features within this model family. It also demonstrates the converse: keeping synaptic strengths fixed while changing intrinsic cellular parameters can produce very different rhythms. It does not show that all parameter combinations work, that connectivity is irrelevant, or that every successful model is equivalent under every condition.

Circuit and the meaning of “fixed topology”

The biological reference is the pyloric motor circuit of the lobster Homarus americanus. Its anterior burster (AB) neuron is electrically coupled to two pyloric dilator (PD) neurons; LP is a single lateral pyloric neuron, while PY comprises five to eight neurons. These cells generate a triphasic burst sequence, conventionally PD–LP–PY.

The model reduces the circuit to three single-compartment units:

• AB and both PD neurons are collapsed into one AB/PD pacemaker.
• One LP model represents LP.
• One PY model represents the PY population.

VD and IC cell types are omitted. Electrical coupling within the pacemaker is absorbed into the lumped AB/PD unit. No detailed dendritic morphology, organ, muscle-feedback model or whole-animal behavior is reconstructed.

There are seven allowable directed inhibitory synaptic slots: fast glutamatergic and slow cholinergic connections from AB/PD to each follower, glutamatergic LP feedback to AB/PD, and reciprocal glutamatergic LP–PY connections. The fast/slow labels retain the separately modeled outputs of biologically distinct AB and PD cells despite their shared model voltage.

The permissible connection architecture and synaptic classes are fixed, but calling the entire search strictly fixed nonzero topology would be inaccurate: 0 nS is included, effectively removing a connection. The authors explicitly discuss resulting open networks. Figure 3f–j nevertheless provides a stricter comparison: the same complete synaptic-strength vector gives different network rhythms when the chosen intrinsic neuron models change. Thus the cellular-parameter result does not depend solely on allowing edge removal.

What varies and what is held fixed

Each model neuron contains eight currents: sodium, fast and slow transient calcium, transient potassium, calcium-dependent potassium, delayed-rectifier potassium, hyperpolarization-activated inward current, and leak. Voltage dependence and channel dynamics are inherited from an earlier STG model and held identical across neuron choices; maximal conductance densities differ. The paper supplies all eight conductances for each of the 16 selected neuron models in Table 2.

Neuron parameters are not sampled as arbitrary independent continuous values in this study. They occur in preselected bundles: five AB/PD, five LP and six PY models, giving 150 cell combinations. The bundles were selected from an earlier approximately 1.7-million-neuron database using substantial biological-response filtering:

• AB/PD candidates had periods of 1–2 seconds, burst durations .5–.75 seconds, duty cycles .3–.4, and suitable phase-response curves/slow-wave amplitudes. Nine qualified; five were selected to span intrinsic properties. Their isolated periods were 1.46, 1.49, 1.58, 1.61 and 1.64 seconds.
• Follower candidates were silent or slowly spiking below 11.5 Hz, with additional resting/minimum-voltage requirements. They had to rebound to both weak and strong pacemaker-derived inhibitory inputs with at least five spikes over at least 200 ms before an interspike interval exceeding 100 ms.
• LP rebound delay after weak inhibition had to be below 250 ms; PY delay had to be 300–500 ms. Five of 31 eligible LP models and six of 57 eligible PY models were chosen to span rebound/intrinsic properties.

All seven synaptic strengths are then varied independently over a finite grid: 0, 3, 10, 30, 100 nS, with 1 nS additionally included for the three synapses onto PY. That makes 135,000 synaptic configurations × 150 cell combinations = 20,250,000 networks. The authors describe 100 nS as sufficiently strong to approximately clamp the postsynaptic cell during input. This chosen grid is not a probability distribution over biological animals or the full continuous physiological parameter space.

Synaptic current depends on maximal conductance, activation and driving force. Activation follows a voltage-dependent first-order kinetic model. Glutamatergic synapses use reversal potential −70 mV and dissociation rate 1/40 ms; cholinergic synapses use −80 mV and 1/100 ms. Both share half-activation −35 mV and slope parameter 5 mV. These kinetics and reversal potentials are fixed assumptions, not varied uncertainties or inferred receptor measurements. “Synaptic strength” here means a modeled conductance, not an anatomical count of synaptic contacts.

Simulation and acceptance criteria

Each neuron starts from a saved point on its isolated dynamics and synapses start deactivated. After three seconds allowed for transients, the program detects spikes and attempts classification in successive one-second epochs, stopping when it successfully classifies the circuit. The implementation uses C++ and exponential Euler integration. The article does not provide a systematic multiple-initial-condition/basin analysis, long-horizon robustness validation or a numerical step-size convergence analysis. Important implementation detail is inherited from referenced work and was not independently verified here.

Three levels of output restriction should remain separate:

The empirical limits derive from control recordings in 99 lobster preparations, described here as the authors' unpublished recording collection. Those animals define rhythm ranges; the paper does not measure all their cellular and synaptic parameters to demonstrate biological degeneracy directly.

The 15 features are cycle period; three burst durations; two temporal gaps; two delays from PD burst onset; three duty cycles; two phase gaps; and LP/PY onset phases. Burst boundaries are identified using the first and last spike peaks. Table 1 gives all empirical means, SDs and acceptance intervals; period, for example, is 1.509±.279 seconds, with accepted interval .952–2.067 seconds.

Several features are algebraically related, such as duration and duty cycle. Passing 15 criteria is not passing 15 independent experiments. Independent marginal mean±2 SD limits also do not reproduce the biological joint feature distribution. The close-match sets are defined relative to three chosen exemplars, not by exhaustive pairwise equivalence among all accepted networks.

Most importantly, similarity concerns these burst-level timing features. It does not require equality of every action-potential time, number of spikes, within-burst firing rate, voltage trajectory, channel current, energy cost or response to every input. Figure 5 shows similar burst organization alongside visible differences in spike patterns. The abstract's “virtually indistinguishable” should be interpreted within the declared output resolution.

Main results and what remains constrained

All 150 cell combinations appear among both the pyloric-like and the stricter pyloric networks. All selected LP and PY identities also appear in each fast/medium/slow subset, while the pacemaker identities differ between subsets. Network period remains strongly constrained by the chosen pacemaker's intrinsic period: Figure 4 shows prominent peaks associated with those intrinsic periods, with accepted rhythms equal to or slower than the pacemaker.

Most synapse-strength distributions span the full tested range even after rhythm filtering. The major exception is LP→PY: it exceeds 3 nS in only 0.1% of pyloric networks and never exceeds 30 nS. Other connections show preferences rather than complete freedom: strong fast/weak slow inhibition toward LP and relatively strong slow inhibition toward PY favor the correct rebound order; PY→LP tends to be strong. These preferences are compared with prior biological findings in the Discussion.

Figure 5 presents two circuits with similar rhythms despite at least threefold differences in the conductances displayed there. The figure shows selected membrane/synaptic parameters; it would be wrong to infer that every parameter in the full two models differs threefold.

Broad marginal distributions do not mean any parameter can be changed independently without consequence. Accepted networks contain particular combinations; compensation and constraints matter. Nor do the enumerated success fractions estimate how common these configurations are in animals. They are conditional on the selected cell pools, synaptic grid, fixed kinetics, initialization and acceptance test.

Does the paper demonstrate divergence under perturbation?

Not for baseline-matched complete networks in the sense needed for a strong out-of-distribution claim. It changes component and synaptic parameters across the database, and Figure 2 explicitly compares follower-neuron responses to different inhibitory inputs. Figure 3 shows different outputs under different circuit parameters. Those are genuine parameter/input sensitivity demonstrations.

However, the paper does not take the accepted near-matching network ensemble, administer an identical subsequent neuromodulator, lesion or environmental challenge to every member, and measure divergence of their responses. Its Discussion explicitly proposes that requiring adequate performance under multiple neuromodulatory conditions could further restrict the viable models. That is a motivated additional test, not an experiment completed here. Likewise, compensation/homeostasis is offered as an explanation and prediction; no homeostatic learning rule evolves these enumerated networks in this study.

Implications for connectomics and individual models

Within this model family, a synaptic conductance map alone does not uniquely determine the output when intrinsic membrane parameters remain unknown, and a bounded baseline output does not uniquely identify those parameters. This is stronger and more precise than the slogan “the connectome tells us nothing.” The model's wiring, transmitter classes, kinetics and cell-response priors sharply restrict what can be generated; some connections and pacemaker properties remain important.

The result also cuts against an indiscriminate demand for exact recovery of every parameter for every task: many parameter sets suffice for this particular rhythmic output within its tolerance. Whether a particular parameter is necessary depends on the target prediction, perturbations, tolerance and model class. The paper establishes neither that all microscopic details must be preserved nor that a small generic parameter set suffices for arbitrary individual reconstruction.

For Kurisutina, the relevant inference is partial identifiability: several internal models may fit the same visible behavior. Such fits justify an equivalence class for declared tests, not a recovered unique personal mechanism. Species-typical motor rhythms provide no direct evidence about autobiographical memory, subjective experience or continuity of a copied subject. The study has no consciousness endpoint and does not model a whole nervous system.

A concrete follow-up would retain all acceptable models, distinguish identical nonzero topology from edge deletion, and freeze predictions for a shared held-out perturbation battery. Perturbations would be selected where candidate predictions differ, then assessed against individual biological preparations. This would test how much anatomical, intrinsic and state information is needed to predict that individual's responses. It is a research proposal motivated by the paper, not one of its reported results.