Pith. sign in

REVIEW 4 major objections 8 minor 57 references

Synaptic clustering emerges from learning and supports covariance discrimination

T0 review · 4 major / 8 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Dendrites trained on a covariance task spontaneously form functional synapse clusters, and learned placement is needed for the solution.

desk verdict Clean existence result that dendrites + rewiring grow E/I clusters on a covariance task, but the shuffle does not isolate FSCs from fan-in reallocation. read the letter →

arxiv 2607.24503 v1 pith:VVNM2PMZ submitted 2026-07-27 q-bio.NC

Ilenna Simone Jones , Maceo Richards , Houman Safaai , Elom Amematsro , Bernardo Sabatini This is my paper · ORCID
classification q-bio.NC
keywords functionalsynapseclustersdendriticcomputationcovarianceclassificationstructuralplasticityshuntinginhibitionNMDAnonlinearityexcitatory-inhibitoryorganization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Functional synapse clusters—groups of co-tuned inputs sitting on the same dendritic branch—appear after learning in real cortex and hippocampus, but experiments that block them also block the dendritic nonlinearities that may be doing the computation, so necessity has stayed unresolved. This paper builds Dendrinet, a sparse tree of dendritic compartments with conductance-based synapses, and trains it on a four-class task whose only discriminative signal lives in permuted block correlations among inputs. When shunting inhibition, an NMDA-like branch nonlinearity, and a structural rewiring rule are all active, the network solves the task and grows class-selective excitatory and inhibitory clusters on distal branches. Turning any of those pieces off collapses performance and changes clustering; shuffling the learned placement of synapses after training also collapses performance, with inhibitory placement mattering more than excitatory. The result is an in-silico isolation argument: compartmental nonlinearities plus learned organization can support covariance discrimination, and clusters emerge as part of that solution rather than as a side effect.

What carries the argument

Dendrinet on the PCC task: a binary-tree network of conductance-based branches composing shunting inhibition with a piecewise linear-to-tanh (Poirazi) nonlinearity, trained by gradient descent plus DeepST rewiring under a fixed sparse synapse budget; clustering is scored by fan-in-normalized block entropy, and necessity is probed by component ablations and post-training weight/mask shuffles.

What would settle it

A placement shuffle that randomizes only within-block co-localization while holding fan-in, channel reuse, weight distributions, and E–I pairing fixed, yet leaves test accuracy intact—or dual-color imaging under NMDAR block that fails to show the predicted rise in inhibitory clustering with loss of excitatory clusters.

Watch

Extended reading notes

Core claim

Neurons with hierarchical dendrites solve the Permuted-Covariance Classification task and develop class-selective excitatory and inhibitory functional synapse clusters only when both dendritic nonlinearities (shunting inhibition and the Poirazi NMDA-proxy) and structural plasticity are active. Ablating the nonlinearities reduces excitatory clusters and performance while increasing inhibitory clusters; shuffling learned connectivity with nonlinearities held fixed also reduces performance, with inhibitory organization proving more critical than excitatory.

Load-bearing premise

That post-training shuffles and the joint need for nonlinearities plus rewiring isolate a causal role for functional clusters themselves, rather than for broader learned placement (fan-in, channel repetition, and excitation–inhibition pairing) of which clusters are only one part.

Editorial extensions

If this is right

  • NMDAR-block deficits should be read as joint loss of nonlinear computation and reorganization of inhibitory clustering, not only loss of excitatory clusters.
  • Connectomic analyses that treat synapses as neuron-to-neuron edges without dendritic location discard a functionally load-bearing variable.
  • Inhibitory FSCs on distal branches are a predicted motif for covariance-sensitive circuits and should be looked for alongside excitatory ones.
  • Feedforward shared-source E/I onto distal dendrites (as in hippocampal temporoammonic–neurogliaform motifs) is a natural biological analogue of the learned architecture.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A pure FSC knockout that preserves other connectivity statistics would be the decisive next simulation; until then the paper shows necessity of learned placement more cleanly than of clusters alone.
  • The stronger sensitivity to inhibitory shuffles suggests future theory should treat distal E–I co-placement, not excitatory clustering alone, as the computational primitive.
  • Bio-plausible local plasticity rules that still recover the same distal E/I organization would close the gap between task-optimized existence proofs and learnability in tissue.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The manuscript introduces Dendrinet, a feedforward network of four readout neurons, each with a binary-tree dendritic structure, conductance-based (shunting) inhibition, a Poirazi/piecewise-linear-to-tanh NMDAR-proxy nonlinearity, and a structural plasticity rule (DeepST, adapted from DEEP R) that rewires synapses under a fixed budget. The authors train the model on a purpose-built Permuted-Covariance Classification (PCC) task in which class information resides exclusively in permuted block covariance structure, so above-chance performance requires second-order computation (LDA fails, QDA is Bayes-optimal). A full 2^3 ablation (shunting, Poirazi, DeepST) over 40 seeds shows all three components are jointly needed for high accuracy; under the ALL-ON condition, excitatory and inhibitory synapses cluster onto task-defined covariance blocks in distal branches (quantified by a fan-in-corrected normalized entropy metric), and a class-dependent excitation/inhibition correlation mechanism emerges in which non-preferred classes are shunted subthreshold while preferred-class residuals are amplified. Post-training shuffles of learned weights or connectivity degrade performance (inhibitory shuffles more than excitatory). An appendix theorem shows, in a linearized single-branch SNR model, that concentrating synapses in one block maximizes an expected-competitor SNR. The authors frame the work as isolating FSC contributions from the NMDAR-block confound that besets APV experiments, and —to

Significance. If the results hold, this is a useful contribution on two fronts. First, it provides a clean "glass-box" demonstration that FSC-like organization can emerge from task optimization rather than being asserted a priori, with a full 2^3 ablation, 40 seeds per condition, and parameter/depth/sparsity-matched MLP baselines — a level of control uncommon in this literature. Second, the inhibitory findings (distal inhibitory FSCs; increased inhibitory clustering when the NMDAR-proxy is removed) are genuinely novel and generate a falsifiable experimental prediction (redistribution, not abolition, of clustering under NMDAR block) testable with dual-color imaging. The mechanism analysis in Figs. 5-6 (class-dependent E/I correlation placing non-preferred classes subthreshold and preferred classes in the amplifying regime of the Poirazi function) is a concrete, interpretable account, and the appendix supplies a worked SNR theorem rather than a plausibility argument. The authors' explicit hedging ("suggest, but do not isolate") is appropriate; the residual gap between that hedge and the title/§3.5 framing is what the requested control would close.

major comments (4)
  1. [§3.5, Fig. 7; §2.6.5] The mask shuffle reverses two things at once: within-branch block clustering (H_norm, Fig. 4A) and the learned per-branch fan-in allocation (Fig. 3A shows DeepST concentrates synapses on distal branches). Since §3.4/Fig. 5 locate the class-selective E/I correlation mechanism in the distal branches, flattening fan-in alone would dismantle the computation even if within-branch clustering were irrelevant. The authors acknowledge the shuffle 'does not change only FSCs,' but §3.5 is still titled 'Ablation of Functional Synapse Clusters,' and the result is offered as evidence about clustering. A fan-in-preserving control is needed and is straightforward: permute channel identities of synapses within each branch (holding per-branch fan-in and the weight multiset fixed), and separately permute synapses across branches while preserving within-branch block composition. Only the comparison of these
  2. [§2.2.2; Fig. 2F; §3.1] The Poirazi slope m and threshold b were 'used values that yielded optimal performance for a 7-node Dendrinet model with all biological components active.' The headline claim that 'all three biological properties are necessary' (Fig. 2F) therefore rests on comparisons in which the nonlinearity-ablated conditions run at an operating point tuned for the ALL-ON model. The Poirazi-OFF drop could partly reflect a mismatched operating point rather than loss of supralinear amplification per se (e.g., the threshold b is also used to set the initialization operating point, Eq. 6). A sensitivity sweep over (m, b) per condition, or at minimum a re-tuning of remaining hyperparameters for the ablated conditions, is needed to support the necessity claim as stated.
  3. [§2.6.5 (Synapse Shuffle Ablation)] The randomization procedure is underspecified: 'each shuffle is defined by which subset it preserves vs shuffles' does not state the permutation scope (within branch, within neuron, or network-wide), whether E and I synapses are shuffled jointly or within valence pools, whether the one-synapse-per-channel-per-branch constraint is maintained post-shuffle, or whether shuffled masks are rebalanced. These details determine what the control actually controls for and must be stated precisely, ideally with pseudocode.
  4. [Theorem 3.1 and Appendix A] The theorem is honestly caveated, but its assumptions (binary weights, single linearized leaf, expected competitor under uniform permutations, and the z-space Pearson covariance of Eq. 11) are some distance from the trained model (softplus weights, hierarchical tree, and inputs x = Φ(z) whose copula-transformed correlations differ from the stated r=0.9/0.3). Please state explicitly, near the theorem, which of the empirical findings it does and does not explain — in particular that the E/I correlation mechanism of Fig. 5 is only addressed by Theorem A.6, which compares two placements rather than proving optimality — and note that the quantitative ρ values do not carry through the CDF transform.
minor comments (8)
  1. [Figs. 2F, 4, 7] Statistical reporting: N=40 seeds is stated in captions, but no error bars, confidence intervals, or test procedures (which test, correction for multiple comparisons across the 8 conditions) are described for Figs. 2F, 4, or 7. Please add these.
  2. [§2.5, 'Per seed data independence'] Model and dataset seeds are set equal per run (§2.5), so initialization and data realization covary across seeds. This is a minor design concern; consider decorrelating them or noting why this coupling is harmless.
  3. [§2.6.4 / §3.3] The FSC operationalization uses the task's ground-truth blocks as the definition of 'correlated activity.' Since FSCs are empirically defined by observed presynaptic correlations, please state explicitly that here correlation is inferred from the generative block structure of the copula, and note that the transform to x-space preserves block dependence but not exact Pearson values.
  4. [§2.6.4, Eqs. (8)-(10)] The small constant ε in Eqs. (8)-(9) is never given a value; its size can matter for low-weight blocks. Also, the Monte Carlo null (M=2,000) reassigns the branch's observed weights to random locations — please confirm the null is recomputed per branch and per class as the text implies.
  5. [Throughout] Typos and grammar: 'fromation' (Discussion, para. 1); 'nonlinar dendritic compartmnets' (Discussion); 'must to come' (Discussion); 'biologicaly' (Discussion); 'each of the 3 biological components are active' (§2.4); 'set to 1e-6 in for the analyses' (§2.3).
  6. [Fig. 7 caption; §3.3] 'DON'/'DOFF' appear only in the Fig. 7 caption and are never defined in the main text; the SON/PON/DON toggling notation used in §3.3 should be introduced once, in §2.4, and used consistently.
  7. [§3.3, Fig. 4C] The inhibitory-compensation finding (Poirazi-OFF increases inhibitory clustering) is one of the most novel results and connects to the falsifiable prediction in the Discussion; it deserves a quantitative treatment (effect size across seeds) rather than the current qualitative description of Fig. 4C.
  8. [§2.2.1, Eq. (1)] In Eq. (1), the summation index i over upstream branch inputs g_{v,i} V_{out,i} is undefined in the text, and the reuse of 'v' for both valence (Eq. 3) and branch weights is confusing; please disambiguate the notation.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity: FSCs are an emergent measured outcome under external task labels, not defined as the training objective; only mild hyperparameter selection for the ALL-ON condition.

  1. fitted input called prediction [§2.2.2 Nonlinearities (Poirazi hyperparameters)]
    "To choose the hyperparameters m and b, we used values that yielded optimal performance for a 7-node Dendrinet model with all biological components active."

    m and b are selected to maximize performance of the full (ALL-ON) model that is later treated as the reference condition in which FSCs and high accuracy co-occur. Ablation comparisons therefore use a nonlinearity already tuned under the joint presence of shunting, Poirazi, and DeepST. This is weak circularity-adjacent model selection: it does not make FSC emergence or accuracy tautological, but it slightly favors the ALL-ON operating point relative to identity/ablated variants that were not given the same hyperparameter search.

full rationale

The derivation chain is self-contained against external benchmarks. Class labels and block-covariance structure of the PCC task are fixed before training and are independent of Dendrinet parameters. Performance is cross-entropy / held-out accuracy, compared to LDA, QDA, and parameter-matched MLPs. Functional synapse clustering is quantified post hoc by normalized block-entropy against a fan-in-matched Monte Carlo null (Eqs. 7–10), not optimized as a loss term, so emergence of H_norm < 1 under ALL-ON is an empirical result rather than true by construction. Theorem 3.1 derives an SNR advantage for block-concentrated binary synapses from the same covariance model the task uses; that is a supporting calculation for why clustering can help, not a redefinition of the empirical metric or of accuracy. Ablations (shunting / Poirazi / DeepST off) and post-hoc weight/mask shuffles are interventional tests, and the paper itself states that the shuffle “does not change only FSCs” and only “suggests, but do[es] not isolate” an FSC contribution—so the central claim is not forced by definition. The sole mild circularity-adjacent choice is selecting Poirazi (m, b) for best ALL-ON 7-node performance before running the ablation suite; that is ordinary model selection, not a fitted quantity renamed as an independent prediction. Self-citations (e.g. Jones & Kording dendritic ML work) are background, not uniqueness theorems that forbid alternatives. Score 1 reflects that minor hyperparameter dependence only.

Assumptions & free parameters 6 free parameters · 6 assumptions · 4 invented entities

The central existence claim rests on a constructed task, a hand-specified dendritic activation stack, a non-biological optimizer with a fixed synapse budget, and several tuned scalars (Poirazi m,b; lrs; prune threshold). Biological interpretation further assumes the Poirazi piece-wise map is an adequate NMDAR proxy and that feedforward shared-source E/I onto a binary tree captures the relevant circuit motif. No new physical entity is postulated; Dendrinet/DeepST/PCC/H_norm are methodological constructs.

free parameters (6)
  • Poirazi slope m and threshold b = selected for best ALL-ON 7-node performance (numeric values not reported in text)
    Chosen as values that yielded optimal performance for the full 7-node model with all components on; ablations inherit this tuning.
  • DeepST prune weight threshold = 1e-6
    Synapses with effective weight below threshold are dropped each step; sets effective plasticity aggressiveness.
  • Fixed synapse budget per neuron (E and I) = 140 excitatory and 140 inhibitory synapses per neuron
    Active count held fixed while locations move; shapes fan-in and clustering capacity.
  • Synapse and branch-weight learning rates / weight decay = synapse lr=0.1; branch lr=0.001; weight decay=1e-2; grad clip=5.0
    Adam group lrs and decay are hand-set and material to whether DeepST and weights co-adapt.
  • PCC correlation levels and block geometry = r_in=0.9, r_out=0.3; N=200; 20×10 blocks; 4 classes
    Within-block r=0.9, across-block r=0.3, 20 blocks of 10 in 200-D define task difficulty and the clustering null.
  • Initial sparsity / target fan-in scale = average 20 synapses per branch at initialization
    ~10% channels per branch at init (avg 20 synapses) under neuron-wide budget influences learnable organization.
assumptions (6)
  • domain assumption Steady-state conductance equation with E rev=1, I/leak rev=0 yields divisive shunting (Eq. 1).
    Standard reduced biophysics (Koch); ignores dynamics, driving-force changes, and morphology beyond discrete nodes.
  • domain assumption Poirazi piecewise linear-to-tanh is a valid computational proxy for NMDAR dendritic spikes (Eq. 2).
    Taken from Poirazi/Polsky line; used as on/off ablation stand-in for APV experiments.
  • ad hoc to paper Class information in PCC lives only in second-order block structure after matched marginals (Gaussian copula + CDF).
    Task is purpose-built so linear first-order nets fail and clustering is a plausible solution.
  • ad hoc to paper Backprop + DeepST rewiring is an acceptable optimizer to ask whether FSC-like organization can support the computation (existence), not how biology learns it.
    Explicitly acknowledged in Discussion; load-bearing for interpreting trained connectivity as relevant to brain FSCs.
  • domain assumption Binary-tree discrete branches with at most one synapse per input channel per branch suffice to represent functional clustering.
    Branch-as-subunit abstraction; discards within-branch micron-scale geometry noted as a limitation.
  • standard math For fixed synapse count s≤n, expected-competitor SNR is maximized by placing all synapses in one covariance block (Theorem 3.1).
    Appendix derivation under linearized variance / block-covariance model; motivates why clustering helps but is not the trained network.
invented entities (4)
  • Dendrinet
    purpose: Hierarchical sparse E/I dendritic ANN used as the trainable glass-box neuron/circuit.
    Architecture invented for this study; composed of known biophysical motifs.
  • DeepST (Deep Synaptic Translocation)
    purpose: Prune-and-replace structural plasticity paired with gradient descent under fixed density.
    Named variant of DEEP R-style rewiring adapted to Dendrinet masks.
  • Permuted-Covariance Classification (PCC) task
    purpose: Force discrimination of input covariance block permutations with matched first-order stats.
    Custom benchmark so above-chance performance requires second-order structure.
  • Normalized entropy clustering metric H_norm
    purpose: Fan-in-corrected measure of weight mass concentration on class-specific covariance blocks.
    Defined in Methods via Monte Carlo null; operational definition of FSCs in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synaptic clustering emerges from learning and supports covariance discrimination." pith.science (2026). https://pith.science/paper/VVNM2PMZ

@misc{pith2026260724503,
  author       = {Pith},
  title        = {Pith review of: Synaptic clustering emerges from learning and supports covariance discrimination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VVNM2PMZ}},
  note         = {Machine review of arXiv:2607.24503}
}
read the original abstract

Functional synapse clusters (FSCs) are synapses with correlated presynaptic activity that are colocalized on the same neuronal dendritic branch. FSCs have been observed after learning in cortical and hippocampal pyramidal neurons. However, previous efforts to ablate FSCs by pharmacologically blocking dendritic nonlinearities to establish causal necessity may have confounded effects. Therefore, whether FSCs are causally necessary for computation is unknown. Here, we attempt to isolate FSCs from this potential confounder in silico. We train Dendrinet, an artificial neural network architecture with hierarchical dendritic segments and sparse conductance-based synapses, on a Permuted-Covariance Classification (PCC) task. This task cannot be solved by single-layer linear-nonlinear artificial neural networks. We find that neurons with dendrites can be trained to solve the task and develop excitatory and inhibitory FSCs if both dendritic nonlinearities and synaptic structural plasticity are active. Turning off dendritic nonlinearities reduces excitatory FSCs, which replicates experimental findings, and reduces performance while unexpectedly increasing inhibitory FSCs. Furthermore, shuffling learned synaptic connectivity while keeping the nonlinearities fixed reduces performance. This shows sensitivity to learned connectivity, but the shuffle does not change only FSCs. Shuffling inhibitory synapse properties reduces performance more than the corresponding excitatory shuffle, showing higher sensitivity to inhibitory organization. This work suggests that dendritic compartmentalization and learned synaptic organization can support computation of covariance structure.

Figures

Figures reproduced from arXiv: 2607.24503 by the authors.

Figure 1
Figure 1. Task optimization framework for dendritic network model a. Permuted Covariance Classification (PCC) Task. Each class differs by permutation of input channels, which interchanges the rows and columns of their correlation matrices. Every channel belongs to a "block" of correlated input dimensions. No single channel has discrimination information because the first-order statistics (mean and variance) are equal across a… view at source ↗
Figure 2
Figure 2. Performance and parameter efficiency in Dendrinet compared to simple MLP. a. Multilayer perceptrons (MLPs) tested on PCC, including a 1-hidden layer and 2-hidden layer dense MLP and 1-hidden layer sparse MLP. All MLPs are parameter-size matched to the Dendrinet models. b. Dendrinet models with binary tree structure with increasing depth. 1 node Dendrinet is similar to a zero-hidden layer sparse MLP. Each Dendrinet h… view at source ↗
Figure 3
Figure 3. Dendrinet with all biological properties active exhibits learned connectivity after task optimization a. Distribution of synapses per branch, grouped by morphological branch type. Excitatory and inhibitory organization of synaptic branch targeting sees the most change when all three properties (shunting inhibition, Poirazi nonlinearity, and DeepST) are active. Distal branches have the most excitatory and inhibitory … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Normalized entropy as a metric of functional synapse clustering reveals different excitatory clustering signatures. Normalized entropy yields a task-relevant organization of synaptic weights into functional synapse clusters for both distal and proximal excitatory synap…
Figure 5
Figure 5. Figure 5: Dendrinet branch-targeted synapse correlation, shunting inhibition activation, and branch output. a. Dendrinet branch activity with all three selectable biological properties active. Encoding of class specific information is mediated by excitatory and inhibitory correl…
Figure 6
Figure 6. Figure 6: Correlation between excitation and inhibition precisely allows Poirazi nonlinearity to amplify class-selective signals. a. Variance of the distribution of shunting inhibition output in the Poirazi-ON case in the distal branches are small for non-preferred classes and l…
Figure 7
Figure 7. Figure 7: Shuffle ablation of excitatory and inhibitory synapse clusters significantly reduces performance. Shuffling synaptic components (weight and connectivity) disrupts task performance across the dendritic network. Excitatory and inhibitory synapses are shuffled alone, then…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 26 canonical work pages

  1. [1]

    Florencia Iacaruso, Ioana T

    M. Florencia Iacaruso, Ioana T. Gasler, and Sonja B. Hofer. Synaptic organization of visual space in primary visual cortex.Nature, 547(7664):449–452, July 2017. ISSN 1476-4687. doi: 10.1038/nature23019

    work page doi:10.1038/nature23019 2017

  2. [2]

    Synaptic Clustering and Memory Formation.Front

    George Kastellakis and Panayiota Poirazi. Synaptic Clustering and Memory Formation.Front. Mol. Neurosci., 12, December 2019. ISSN 1662-5099. doi: 10.3389/fnmol.2019.00300

    arXiv 2019

  3. [3]

    Dendritic Computation.Annual Review of Neuroscience, 28(1): 503–532, 2005

    Michael London and Michael Häusser. Dendritic Computation.Annual Review of Neuroscience, 28(1): 503–532, 2005. doi: 10.1146/annurev.neuro.28.061604.135703

    arXiv 2005

  4. [4]

    Stuart and Nelson Spruston

    Greg J. Stuart and Nelson Spruston. Dendritic integration: 60 years of progress.Nat Neurosci, 18(12): 1713–1721, December 2015. ISSN 1546-1726. doi: 10.1038/nn.4157

    doi:10.1038/nn.4157 2015

  5. [5]

    Wilson, David E

    Daniel E. Wilson, David E. Whitney, Benjamin Scholl, and David Fitzpatrick. Orientation selectivity and the functional clustering of synaptic inputs in primary visual cortex.Nat Neurosci, 19(8):1003–1009, August 2016. ISSN 1097-6256. doi: 10.1038/nn.4323

    work page doi:10.1038/nn.4323 2016

  6. [6]

    Ami Citri and Robert C. Malenka. Synaptic Plasticity: Multiple Forms, Functions, and Mechanisms. Neuropsychopharmacol, 33(1):18–41, January 2008. ISSN 1740-634X. doi: 10.1038/sj.npp.1301559

    doi:10.1038/sj.npp.1301559 2008

  7. [7]

    Experience-dependent structural synaptic plasticity in the mam- malian brain.Nat Rev Neurosci, 10(9):647–658, September 2009

    Anthony Holtmaat and Karel Svoboda. Experience-dependent structural synaptic plasticity in the mam- malian brain.Nat Rev Neurosci, 10(9):647–658, September 2009. ISSN 1471-0048. doi: 10.1038/nrn2699

    doi:10.1038/nrn2699 2009

  8. [8]

    Activity-Dependent Clustering of Functional Synaptic Inputs on Developing Hippocampal Den- drites.Neuron, 72(6):1012–1024, December 2011

    Thomas Kleindienst, Johan Winnubst, Claudia Roth-Alpermann, Tobias Bonhoeffer, and Christian Lohmann. Activity-Dependent Clustering of Functional Synaptic Inputs on Developing Hippocampal Den- drites.Neuron, 72(6):1012–1024, December 2011. ISSN 0896-6273. doi: 10.1016/j.neuron.2011.10.015

    work page doi:10.1016/j.neuron.2011.10.015 2011

Show all 57 references
  1. [9]

    Locally Synchronized Synaptic Inputs.Science, 335(6066):353–356, January 2012

    Naoya Takahashi, Kazuo Kitamura, Naoki Matsuo, Mark Mayford, Masanobu Kano, Norio Matsuki, and Yuji Ikegaya. Locally Synchronized Synaptic Inputs.Science, 335(6066):353–356, January 2012. doi: 10.1126/science.1210362

    2012 doi

  2. [10]

    The dendritic engram.Front Behav Neurosci, 17:1212139, July 2023

    George Kastellakis, Simone Tasciotti, Ioanna Pandi, and Panayiota Poirazi. The dendritic engram.Front Behav Neurosci, 17:1212139, July 2023. ISSN 1662-5153. doi: 10.3389/fnbeh.2023.1212139

    2023

  3. [11]

    Frank, Shan Huang, Miou Zhou, Amos Gdalyahu, George Kastellakis, Tawnie K

    Adam C. Frank, Shan Huang, Miou Zhou, Amos Gdalyahu, George Kastellakis, Tawnie K. Silva, Elaine Lu, Ximiao Wen, Panayiota Poirazi, Joshua T. Trachtenberg, and Alcino J. Silva. Hotspots of dendritic spine turnover facilitate clustered spine addition and learning and memory.Nat...

  4. [12]

    Fear extinction reverses dendritic spine formation induced by fear conditioning in the mouse auditory cortex.Proc Natl Acad Sci U S A, 115(37):9306–9311, September 2018

    Cora Sau Wan Lai, Avital Adler, and Wen-Biao Gan. Fear extinction reverses dendritic spine formation induced by fear conditioning in the mouse auditory cortex.Proc Natl Acad Sci U S A, 115(37):9306–9311, September 2018. ISSN 0027-8424. doi: 10.1073/pnas.1801504115

    2018 doi

  5. [13]

    Bloss, Mark S

    Erik B. Bloss, Mark S. Cembrowski, Bill Karsh, Jennifer Colonell, Richard D. Fetter, and Nelson Spruston. Single excitatory axons form clustered synapses onto CA1 pyramidal cell dendrites.Nat Neurosci, 21(3): 353–363, March 2018. ISSN 1546-1726. doi: 10.1038/s41593-018-0084-6

    2018 doi

  6. [14]

    M. L. Mayer, G. L. Westbrook, and P. B. Guthrie. V oltage-dependent block by Mg2+ of NMDA responses in spinal cord neurones.Nature, 309(5965):261–263, May 1984. ISSN 0028-0836. doi: 10.1038/309261a0

    1984 doi

  7. [15]

    Nowak, P

    L. Nowak, P. Bregestovski, P. Ascher, A. Herbet, and A. Prochiantz. Magnesium gates glutamate-activated channels in mouse central neurones.Nature, 307(5950):462–465, February 1984. ISSN 0028-0836. doi: 10.1038/307462a0

    1984 doi

  8. [16]

    Koester, and Yitzhak Schiller

    Jackie Schiller, Guy Major, Helmut J. Koester, and Yitzhak Schiller. NMDA spikes in basal dendrites of cortical pyramidal neurons.Nature, 404(6775):285–289, March 2000. ISSN 1476-4687. doi: 10.1038/ 35005094

  9. [17]

    Larkum, and Jackie Schiller

    Guy Major, Matthew E. Larkum, and Jackie Schiller. Active Properties of Neocortical Pyramidal Neuron Dendrites.Annual Review of Neuroscience, 36(V olume 36, 2013):1–24, July 2013. ISSN 0147-006X, 1545-4126. doi: 10.1146/annurev-neuro-062111-150343. 17

    2013 doi

  10. [18]

    R. G. M. Morris, E. Anderson, G. S. Lynch, and M. Baudry. Selective impairment of learning and blockade of long-term potentiation by an N-methyl-D-aspartate receptor antagonist, AP5.Nature, 319(6056): 774–776, February 1986. ISSN 1476-4687. doi: 10.1038/319774a0

    1986 doi

  11. [19]

    Trachtenberg, and Alcino J

    Megha Sehgal, Daniel Almeida Filho, George Kastellakis, Sungsoo Kim, Jinsu Lee, Yang Shen, Shan Huang, Ayal Lavi, Giselle Fernandes, Irene Davila Mejia, Sunaina Soans Martin, Asli Pekcan, Melody Shana Wu, Won Do Heo, Panayiota Poirazi, Joshua T. Trachtenberg, and Alcino J. Sil...

    2025 doi

  12. [20]

    Antic, Wen-Liang Zhou, Anna R

    Srdjan D. Antic, Wen-Liang Zhou, Anna R. Moore, Shaina M. Short, and Katerina D. Ikonomu. The decade of the dendritic NMDA spike.Journal of Neuroscience Research, 88(14):2991–3001, 2010. ISSN 1097-4547. doi: 10.1002/jnr.22444. _eprint: https://onlinelibrary.wiley.com/doi/pdf/1...

    2010 doi

  13. [21]

    Panayiota Poirazi, Terrence Brannon, and Bartlett W. Mel. Pyramidal Neuron as Two-Layer Neural Network.Neuron, 37(6):989–999, March 2003. ISSN 0896-6273. doi: 10.1016/S0896-6273(03)00149-1

    2003 doi

  14. [22]

    Mel, and Jackie Schiller

    Alon Polsky, Bartlett W. Mel, and Jackie Schiller. Computational subunits in thin dendrites of pyramidal cells.Nat Neurosci, 7(6):621–627, June 2004. ISSN 1097-6256. doi: 10.1038/nn1253

    2004 doi

  15. [23]

    Single cortical neurons as deep artificial neural networks.Neuron, 109(17):2727–2739.e3, September 2021

    David Beniaguev, Idan Segev, and Michael London. Single cortical neurons as deep artificial neural networks.Neuron, 109(17):2727–2739.e3, September 2021. ISSN 1097-4199. doi: 10.1016/j.neuron.2021. 07.002

    2021 doi

  16. [24]

    What can a neuron compute, June 2026

    Ido Aizenbud, David Beniaguev, Noam Pnueli, Idan Segev, and Michael London. What can a neuron compute, June 2026

    2026

  17. [25]

    Daniel L. K. Yamins and James J. DiCarlo. Using goal-driven deep learning models to understand sensory cortex.Nat Neurosci, 19(3):356–365, March 2016. ISSN 1546-1726. doi: 10.1038/nn.4244

    2016 doi

  18. [26]

    Richards, Timothy P

    Blake A. Richards, Timothy P. Lillicrap, Philippe Beaudoin, Yoshua Bengio, Rafal Bogacz, Amelia Christensen, Claudia Clopath, Rui Ponte Costa, Archy de Berker, Surya Ganguli, Colleen J. Gillon, Danijar Hafner, Adam Kepecs, Nikolaus Kriegeskorte, Peter Latham, Grace W. Lindsay,...

    2019

  19. [27]

    Might a Single Neuron Solve Interesting Machine Learning Problems Through Successive Computations on Its Dendritic Tree?Neural Computation, 33(6):1554–1571, May 2021

    Ilenna Simone Jones and Konrad Paul Kording. Might a Single Neuron Solve Interesting Machine Learning Problems Through Successive Computations on Its Dendritic Tree?Neural Computation, 33(6):1554–1571, May 2021. ISSN 0899-7667. doi: 10.1162/neco_a_01390

    2021 doi

  20. [28]

    Bartlett W. Mel. NMDA-Based Pattern Discrimination in a Modeled Cortical Neuron.Neural Comput, 4 (4):502–517, July 1992. ISSN 0899-7667. doi: 10.1162/neco.1992.4.4.502

    1992 doi

  21. [29]

    B. W. Mel, D. L. Ruderman, and K. A. Archie. Translation-invariant orientation tuning in visual "complex" cells could derive from intradendritic computations.J Neurosci, 18(11):4325–4334, June 1998. ISSN 0270-6474. doi: 10.1523/JNEUROSCI.18-11-04325.1998

    1998

  22. [30]

    Ujfalussy and Judit K

    Balázs B. Ujfalussy and Judit K. Makara. Impact of functional synapse clusters on neuronal response selectivity.Nat Commun, 11(1):1413, March 2020. ISSN 2041-1723. doi: 10.1038/s41467-020-15147-6

    2020 doi

  23. [31]

    Deep Rewiring: Training very sparse deep networks, August 2018

    Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep Rewiring: Training very sparse deep networks, August 2018. arXiv:1711.05136

    2018 arXiv

  24. [32]

    Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence.Cogn Comput, 16(1):45–74, January 2024

    Vikas Hassija, Vinay Chamola, Atmesh Mahapatra, Abhinandan Singal, Divyansh Goel, Kaizhu Huang, Simone Scardapane, Indro Spinelli, Mufti Mahmud, and Amir Hussain. Interpreting Black-Box Models: A Review on Explainable Artificial Intelligence.Cogn Comput, 16(1):45–74, January 2...

    2024 doi

  25. [33]

    Oxford University Press, New York, New York, 1999

    Christoph Koch.Biophysics of Computation: Information Processing in Single Neurons. Oxford University Press, New York, New York, 1999. ISBN 0-19-510491-9

    1999

  26. [34]

    Panayiota Poirazi, Terrence Brannon, and Bartlett W. Mel. Arithmetic of Subthreshold Synaptic Summation in a Model CA1 Pyramidal Cell.Neuron, 37(6):977–987, March 2003. ISSN 0896-6273. doi: 10.1016/ S0896-6273(03)00148-X

    2003

  27. [35]

    Friedman.The Elements of Statistical Learning: Data Mining, Inference, and Prediction

    Trevor Hastie, Robert Tibshirani, and Jerome H. Friedman.The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, 2009. ISBN 978-0-387-84884-6. 18

    2009

  28. [36]

    Neuronal ensembles: Building blocks of neural circuits

    Rafael Yuste, Rosa Cossart, and Emre Yaksi. Neuronal ensembles: Building blocks of neural circuits. Neuron, 112(6):875–892, March 2024. ISSN 08966273. doi: 10.1016/j.neuron.2023.12.008

    2024 doi

  29. [37]

    Robert C. Froemke. Plasticity of Cortical Excitatory-Inhibitory Balance.Annual Review of Neuro- science, 38(V olume 38, 2015):195–219, July 2015. ISSN 0147-006X, 1545-4126. doi: 10.1146/ annurev-neuro-071714-034002. URL https://www.annualreviews.org/content/journals/10. 1146/a...

    2015

  30. [38]

    Perceptron Learning and Classification in a Modeled Cortical Pyramidal Cell.Front Comput Neurosci, 14:33, 2020

    Toviah Moldwin and Idan Segev. Perceptron Learning and Classification in a Modeled Cortical Pyramidal Cell.Front Comput Neurosci, 14:33, 2020. ISSN 1662-5188. doi: 10.3389/fncom.2020.00033

    2020

  31. [39]

    Dendritic action potentials and computation in human layer 2/3 cortical neurons.Science, 367(6473):83–87, January 2020

    Albert Gidon, Timothy Adam Zolnik, Pawel Fidzinski, Felix Bolduan, Athanasia Papoutsi, Panayiota Poirazi, Martin Holtkamp, Imre Vida, and Matthew Evan Larkum. Dendritic action potentials and computation in human layer 2/3 cortical neurons.Science, 367(6473):83–87, January 2020...

    2020

  32. [40]

    Do Biological Constraints Impair Dendritic Computation? Neuroscience, 489:262–274, May 2022

    Ilenna Simone Jones and Konrad Paul Kording. Do Biological Constraints Impair Dendritic Computation? Neuroscience, 489:262–274, May 2022. ISSN 0306-4522. doi: 10.1016/j.neuroscience.2021.07.036

    2022 doi

  33. [41]

    Dendritic Integration Inspired Artificial Neural Networks Capture Data Correlation

    Chongming Liu, Jingyang Ma, Songting Li, and Douglas Zhou. Dendritic Integration Inspired Artificial Neural Networks Capture Data Correlation. InAdvances in Neural Information Pro- cessing Systems, volume 37, pages 79325–79349. Curran Associates, Inc., 2024. doi: 10. 52202/079...

    2024

  34. [42]

    Ujfalussy, Judit K

    Balázs B. Ujfalussy, Judit K. Makara, Máté Lengyel, and Tiago Branco. Global and Multiplexed Dendritic Computations under In Vivo-like Conditions.Neuron, 100(3):579–592.e5, November 2018. ISSN 0896-

    2018

  35. [43]

    Dendrites endow artificial neural networks with accurate, robust and parameter-efficient learning.Nat Commun, 16(1):943, January 2025

    Spyridon Chavlis and Panayiota Poirazi. Dendrites endow artificial neural networks with accurate, robust and parameter-efficient learning.Nat Commun, 16(1):943, January 2025. ISSN 2041-1723. doi: 10.1038/s41467-025-56297-9

    2025 doi

  36. [44]

    Hansen, and Bratislav Misic

    Vincent Bazinet, Justine Y . Hansen, and Bratislav Misic. Towards a biologically annotated brain con- nectome.Nat. Rev. Neurosci., 24(12):747–760, December 2023. ISSN 1471-003X, 1471-0048. doi: 10.1038/s41583-023-00752-3

    2023 doi

  37. [45]

    Alexander Bae, Mahaly Baptiste, Maya R

    J. Alexander Bae, Mahaly Baptiste, Maya R. Baptiste, Caitlyn A. Bishop, Agnes L. Bodor, Derrick Brittain, Victoria Brooks, JoAnn Buchanan, Daniel J. Bumbarger, Manuel A. Castro, Brendan Celii, Erick Cobos, Forrest Collman, Nuno Maçarico da Costa, Bethanny Danskin, Sven Dorkenw...

    2025

  38. [46]

    Vasserman, HyoJong Jang, Bryce W

    Anthony Moreno-Sanchez, Alexander N. Vasserman, HyoJong Jang, Bryce W. Hina, Catherine R. von Reyn, and Jessica Ausborn. Morphology and synapse topography optimize linear encoding of synapse numbers in Drosophila looming responsive descending neurons.eLife, 13, October 2024. d...

    2024 doi

  39. [47]

    Chen, Katherine L

    Jerry L. Chen, Katherine L. Villa, Jae Won Cha, Peter T. C. So, Yoshiyuki Kubota, and Elly Nedivi. Clustered Dynamics of Inhibitory Synapses and Dendritic Spines in the Adult Neocortex.Neuron, 74(2): 361–373, April 2012. ISSN 0896-6273. doi: 10.1016/j.neuron.2012.02.030

    2012 doi

  40. [48]

    Abbott, Idan Segev, Hanchuan Peng, and Franck Polleux

    Daniel Maxim Iascone, Yujie Li, Uygar Sümbül, Michael Doron, Hanbo Chen, Valentine Andreu, Finola Goudy, Heike Blockus, Larry F. Abbott, Idan Segev, Hanchuan Peng, and Franck Polleux. Whole-Neuron Synaptic Mapping Reveals Spatially Precise Excitatory/Inhibitory Balance Limitin...

    2020 doi

  41. [49]

    Poorthuis, Daniella Apelblat, Karzan Muhammad, M

    Elisabeth Abs, Rogier B. Poorthuis, Daniella Apelblat, Karzan Muhammad, M. Belen Pardi, Leona Enke, Dahlia Kushinsky, De-Lin Pu, Max Ferdinand Eizinger, Karl-Klaus Conzelmann, Ivo Spiegel, and Johannes J. Letzkus. Learning-Related Plasticity in Dendrite-Targeting Layer 1 Inter...

    2018 doi

  42. [50]

    Neurogliaform cells and other interneurons of stratum lacunosum-moleculare gate entorhinal-hippocampal dialogue.J Physiol, 589(Pt 8):1875–1883, April 2011

    Marco Capogna. Neurogliaform cells and other interneurons of stratum lacunosum-moleculare gate entorhinal-hippocampal dialogue.J Physiol, 589(Pt 8):1875–1883, April 2011. ISSN 1469-7793. doi: 10.1113/jphysiol.2010.201004

    2011

  43. [51]

    Price, Bruno Cauli, Endre R

    Christopher J. Price, Bruno Cauli, Endre R. Kovacs, Akos Kulik, Bertrand Lambolez, Ryuichi Shigemoto, and Marco Capogna. Neurogliaform Neurons Form a Novel Inhibitory Network in the Hippocampal CA1 Area.J. Neurosci., 25(29):6775–6786, July 2005. ISSN 0270-6474, 1529-2401. doi:...

    2005

  44. [52]

    Matteo Carandini and David J. Heeger. Normalization as a canonical neural computation.Nat Rev Neurosci, 13(1):51–62, January 2012. ISSN 1471-0048. doi: 10.1038/nrn3136

    2012 doi

  45. [53]

    The gradient clusteron: A model neuron that learns to solve classification tasks via dendritic nonlinearities, structural plasticity, and gradient descent

    Toviah Moldwin, Menachem Kalmenson, and Idan Segev. The gradient clusteron: A model neuron that learns to solve classification tasks via dendritic nonlinearities, structural plasticity, and gradient descent. PLOS Computational Biology, 17(5):e1009015, May 2021. ISSN 1553-7358....

    2021 doi

  46. [54]

    Baggetta, Brian M

    Zhe Dong, Yu Feng, Keziah Diego, Austin M. Baggetta, Brian M. Sweis, Zachary T. Pennington, Sophia I. Lamsifer, Yosif Zaki, Federico Sangiuliano, Paul A. Philipsberg, Denisse Morales-Rodriguez, Daniel Kircher, Paul Slesinger, Tristan Shuman, Daniel Aharoni, and Denise J. Cai. ...

  47. [2018]

    doi: 10.1038/s41467-017-02751-2

    ISSN 2041-1723. doi: 10.1038/s41467-017-02751-2

    doi

  48. [2025]

    by construction

    doi: 10.1126/sciadv.adr6470. 20 A Functional Synapse Cluster benefit proof Σk = (1−ρ l)IN + (ρl −ρ g) NbX j Pke(j)e(j)⊤ P ⊤ k +ρ g11⊤,(11) wheree (j) ∈ {0,1}N is a binary vector with nonzero entries at the indices satisfying jN Nb ≤i < (j+1)N Nb . Lemma A.1(Precision matrix de...

    doi

  49. [6273]

    doi: 10.1016/j.neuron.2018.08.032

    2018 doi

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.