acceptodds
Under review as a conference paper at ICLR 2027

Evidence Accumulation in Synaptic Efficacy: Activity-Silent Decision-Making with Short-Term Plasticity

Abstract

Decision-making requires accumulation of evidence over time, yet exactly how neural circuits implement this computation remains debated. Classical models posit that evidence accumulation is realized by ramping neural activity, whereas recent experimental findings suggest an alternative mechanism in which neural activity remains largely silent before abruptly transitioning into an active decision state. This mechanism leaves a fundamental question unanswered: how is evidence accumulated during the neural activity-silent period? To address this question, we trained excitatory–inhibitory recurrent neural networks (RNNs) with short-term synaptic plasticity (STP) on a reaction-time decision task. We found that after training, the networks learned to leverage STP to make decisions in an activity-silent manner. Choice-tuning analysis revealed a clear functional organization of the network, and connection-pruning analysis uncovered the coupling between choice-selective excitatory and inhibitory neural groups. Guided by the architecture of trained RNNs, we constructed a minimal neural circuit model, in which evidence accumulates in facilitated synaptic efficacy that progressively reshapes the circuit's phase portrait, eventually driving the network state across the stability boundary and triggering a rapid transition to a choice state. We further compared activity before a choice between the theoretical STP and Static models to examine the potential energy benefit of activity-silent decision-making. Together, our results suggest STP as a mechanism for evidence accumulation in activity-silent decision-making and offer insights into the design of energy-efficient artificial neural networks.

Primary area:
applications to neuroscience & cognitive science
Keywords:
evidence accumulation, activity-silent decision-making, short-term synaptic plasticity, recurrent neural networks, computational neuroscience

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UHuROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how neural circuits accumulate evidence during activity-silent decision-making. The authors train recurrent neural networks with STP and show that STP promotes step-like decision dynamics. Through circuit analysis and a simplified theoretical model, they propose that evidence can accumulate in synaptic efficacy, eventually triggering a rapid transition to a decision state. The results also suggest potential benefits for energy-efficient neural computation.

Critical strengths
  1. The paper addresses an interesting question that is how can neural circuits accumulate evidence when neuronal activity remains low before a decision.

  2. The comparison with static RNNs shows that STP leads to more step-like decision dynamics, even without explicitly encouraging this behavior during training.

  3. The authors use choice-tuning, connection pruning, and neuronal silencing to identify the neural populations and connections involved in decision-making.

  4. The simplified dynamical model provides a clear explanation for abrupt decision transitions through changes in network stability. The spike-count analysis also suggests potential energy-efficiency benefits.

Critical weaknesses
  1. The central claim that evidence accumulates in synaptic efficacy is not directly demonstrated in the trained RNNs, as the main mechanistic explanation comes from a separately constructed theoretical model.

  2. Step-like activity also occurs in static RNNs, suggesting that STP is not necessary for activity-silent decision-making.

  3. The STP RNNs achieve lower validation accuracy than the static baselines, raising concerns about the trade-off between decision dynamics and task performance.

  4. Silencing high-STP neurons also removes their connections and other computational contributions, making it difficult to isolate the causal role of STP itself.

  5. The energy-efficiency analysis relies on expected spike counts in a simplified model rather than actual energy measurements, and the compared models differ in their decision times.

  6. The experiments focus on a single binary motion-discrimination task with relatively small networks, leaving the generalizability of the proposed mechanism to more complex tasks unclear.

Questions for authors
  1. Can the authors provide more direct evidence that synaptic efficacy is responsible for evidence accumulation in the trained RNNs? For example, what happens to decision accuracy and step-like dynamics if STP is frozen during inference while the learned weights remain unchanged?

  2. How robust is the reported energy-efficiency advantage under a more controlled comparison? Since the STP and static models differ in reaction times and neuronal time constants, would the reduction in expected spike counts still hold if these factors were matched, and synaptic transmission costs were considered?

4 comments
AuthorsOpenReview ↗

Authors reply: Reviewer UHuR (part 1/2)

Thank you again for your insightful review. We appreciate your positive assessment of the central question, circuit analysis, and stability-based explanation. To address your main concerns, we performed same-weight interventions, matched comparisons, and an additional perceptual decision task.

1. Direct evidence linking STP to evidence accumulation (W1, W4, Q1)

Thank you for this insightful question. In the trained RNNs, STP-mediated transmission contributes causally to decision performance and carries a decodable population-level evidence trace. We tested this in ten trained networks by setting effective synaptic efficacy q to one throughout inference while keeping the learned weights unchanged. Separately, we decoded cumulative left-minus-right sensory input from effective synaptic efficacy q, neuronal activity r, and their combination [q,r]:

TestResult
Effective synaptic efficacy set to oneFinal-choice accuracy decreased by 12.21 points overall (95% CI for the paired change, -20.55 to -3.87; p=0.0059) and by 19.40 points on low-coherence trials.
Decode cumulative left-minus-right sensory inputHeld-out R²: neuronal activity r, 0.8930±0.0207; synaptic gain q, 0.9492±0.0136; joint [q,r], 0.9613±0.0098 (mean ± SD across ten networks).

The intervention on STP substantially reduced mean accuracy. Synaptic gain outperformed neuronal activity as a readout in all ten networks and reduced RMSE by 31.2±7.2%. Together, the intervention and decoding results link STP in the trained RNNs to evidence accumulation. We will add these results to the revised manuscript.

2. Activity-silent dynamics in Static RNNs (W2)

Thank you for your comment. A recurrent neural network is a highly expressive nonlinear dynamical system and can, in principle, approximate arbitrary finite-time trajectories, including activity-silent transitions (Funahashi and Nakamura, 1993). We therefore do not claim that STP is necessary. Our evidence instead identifies STP as a strong inductive bias. Across 10 seeds, activity-silent dynamics were more prevalent among choice-tuned neurons in RNNs with STP (52.2±16.2%) than in Static RNNs (36.2±7.0%). Step-like choice-output trajectories on correct-response trials were also more prevalent (58.7±4.0% versus 38.3±3.6%). The fixed-weight intervention and decoding results in our response to W1/Q1 address function: disrupting effective synaptic efficacy impaired decisions, while synaptic gain provided a more accessible readout of accumulated evidence than neuronal activity. Together, these results support STP as an inductive bias for activity-silent, evidence-based decisions, without claiming that it is necessary for step-like activity.

3. Validation accuracy and optimization budget (W3)

Thanks for raising this point. Although Static achieved slightly higher final-choice accuracy on the validation set in the original comparison, this endpoint argmax metric is not the same as the masked trajectory loss used to select checkpoints, nor does it account for timeouts or response timing. It also depends on the STP setting; with U×2 (doubling baseline utilization U), validation accuracy reached 95.14±1.13%, close to the Static value. We therefore also evaluated held-out behavior and decision dynamics across 10 seeds (mean ± SD):

EvaluationSTPStatic
final-choice accuracy95.05±1.17%79.92±6.78%
timeout fraction5.46±2.26%13.96±3.26%
timeout fraction, low coherence (1 and 2)13.55±5.73%30.26±5.62%
Correct threshold crossing within the trial, all trials90.25±1.75%83.44±3.14%

The STP networks performed better on this held-out testing task stream. RNNs with STP had higher final-choice accuracy, fewer timeouts in all ten network pairs, and more correct threshold crossings, especially on weak-evidence trials. This identifies an online benefit for timely, well-separated commitment that final-output accuracy misses. Training also strengthened STP, and the higher step-like fractions indicate that learned STP is an inductive bias toward the activity-silent dynamics. Fixed-weight ablations in response to W1/W4/Q1 above further show that disrupting effective synaptic gain impairs decision performance. Complementary decoding shows that synaptic state carries a more accessible accumulated-evidence trace than neuronal activity. Together, these results support a role for STP in evidence processing and reliable within-trial commitment.

AuthorsOpenReview ↗

Authors reply: Reviewer UHuR (part 2/2)

4. Controlled activity-cost comparison (W5, Q2)

Thanks for your comments. Action potentials and postsynaptic signaling account for major components of neural energy use, so integrated firing rate is a biologically motivated activity proxy (Attwell and Laughlin, 2001; Howarth et al., 2012). Moreover, metabolic costs are difficult to quantify. The variables u and x represent release probability and available synaptic resources, respectively. Neurotransmitter release and vesicle recycling occur at the biological synapses represented by both Static and STP conditions and incur metabolic costs in both (Tsodyks and Wu, 2013); these costs are not quantified here.

Besides, a fair energy comparison should match behavioral performance, particularly accuracy and reaction time, rather than internal parameters like neuronal time constant. Because STP changes circuit dynamics, imposing the same neuronal time constant on the STP and Static models produces different behavior. To compare activity cost at comparable decision performance, we matched accuracy and reaction time while allowing each model's internal parameters to be calibrated separately. This comparison was performed at three behavioral tiers; each entry below is reported as STP / Static:

Target-accuracy tierObserved accuracyMean reaction time (ms)Expected spikes per trial; reduction with STP (95% CI)
85%84.24% / 83.80%264.9 / 266.64.723 / 6.567; 28.08% (27.72%–28.46%)
90%89.01% / 88.70%410.2 / 414.87.675 / 10.269; 25.26% (24.84%–25.70%)
95%94.45% / 94.83%747.4 / 748.915.160 / 18.472; 17.93% (17.69%–18.17%)

Across the three tiers, accuracy differed by at most 0.44 percentage points and mean reaction time by only 0.19%–1.10%; STP nevertheless required fewer expected decision-period spikes at every tier. Thus, across closely matched behavioral operating points, STP achieved comparable accuracy and decision speed with consistently lower spiking activity. Thank you again for the valuable question and we will add the controlled sweep to the revised manuscript.

5. Generalization beyond the present task (W6)

Thank you for this suggestion. Our study focuses on perceptual decision-making, and we therefore added the variable-stimulus-duration task of Song et al. (2016). The reaction-time and variable-stimulus-duration tasks are two canonical perceptual decision-making paradigms rather than tasks selected for simplicity. In the variable-stimulus-duration task, STP strength increased during training, and activity-silent output dynamics were more prevalent in all three paired networks (STP 0.851±0.025 versus Static 0.645±0.010). Thus, this second paradigm reproduces the central findings that task optimization recruits stronger STP and promotes activity-silent decision dynamics. We will add the results to the revised manuscript.

Thank you again for your insightful review. We hope this answers your questions, and we are happy to clarify anything further.

Thank you for your detailed response. It has addressed most of my concerns. I will maintain my original score.

Thank you for reviewing our response. We are glad that it has addressed most of your concerns. We have now incorporated the additional results and clarifications into the revised manuscript, with the main additions highlighted in red. In particular, the fixed-weight interventions and decoding comparisons provide direct trained-RNN evidence linking STP to decision performance and accumulated sensory input (W1, W4, Q1). The Static comparisons clarify STP’s role as an inductive bias toward activity-silent dynamics (W2), and the held-out evaluation shows fewer timeouts and more correct threshold crossings within the trial (W3). The controlled theoretical-model comparison also shows lower expected spike counts for STP at comparable accuracy and reaction time (W5, Q2). Additionally, we clarified the relationship between the trained RNNs and the theoretical model and improved the figures and explanations.

We believe that the additional experiments have strengthened the paper’s soundness, and the revised figures and explanations have also improved its presentation. We would therefore be grateful if you could reconsider your overall assessment, as well as the soundness and presentation ratings, in light of these improvements. If important concerns remain, we would appreciate knowing which ones most influence your assessment so that we can address them directly. Thank you again for engaging with our work and for giving its scientific contributions careful consideration in your final evaluation.

Reviewer_ueQFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether short-term synaptic plasticity (STP) mechanism can be used in excitatory-inhibitory(E-I) neural network for decision-making tasks. The authors first examine an E-I RNN with dynamic STP weight and compare it with a static baseline. More stepping-like activities are observed in STP version. One step ahead, the authors create a network with four populations and observes a change in network stability induced by synaptic efficacy change. The potential advantage of STP mechanism would be a reduction in spike firing energy use.

Critical strengths
  • The paper builds a model to simulate STP mechanism which shows more spontaneous neural stepping activities than the baseline.
  • The four-population model and analysis of synaptic efficacy provides a possible explanation for how STP contributes to step-like decision dynamics.
Critical weaknesses
  1. The novelty is somewhat limited. Previous work, especially Masse et al. (2019), has already studied STP in trained RNNs and its role in activity-silent information storage. Improvement of this paper from previous work is not clear enough.
  2. Limited evidence is shown whether strong STP neurons is directly connected to evidence accumulation. Silencing neurons with strong STP reduces accuracy, but these neurons are also highly choice selective. This does not demonstrate that their STP dynamics are responsible for storing evidence.
  3. The network stability of the four-population model is examined under a frozen network, not in a true case, a dynamic network. It does not mean same thing happens in the real network.
  4. The practical benefit of STP is unclear. The static RNN achieves higher validation accuracy than one with STP and static RNN is not without stepping at all. Also, lower energy consumption is announced by calculating spike activity. But calculation of synaptic efficacy like u or x is not counted. Therefore, energy efficiency is not strongly supported.
Questions for authors
  1. What is the main new contribution compared with Masse et al. (2019) and previous studies of STP in decision-making networks?
  2. Can the authors directly test whether the trained RNNs undergo the proposed stability transition? For example, can changes in synaptic efficacy predict the timing of the transition in neural activity?
  3. Since Static RNNs achieve higher accuracy and can also produce step-like activity, what specific computational advantage does STP provide? Can the energy comparison be performed under more closely matched conditions?
5 comments
AuthorsOpenReview ↗

Authors reply: Reviewer ueQF (part 1/3)

Thank you for your valuable review. We appreciate your recognition of the activity-silent dynamics and four-population analysis. Below we address your concerns point by point.

1. Contribution beyond prior trained-STP studies (W1, Q1)

We believe this work is novel enough. There are three main reasons:

Different computation. Working memory and decision making are two different cognitive functions. Masse et al. (2019) studied the maintenance and manipulation of information across a blank delay. Our task instead requires continuous evidence integration and commitment. The underlying unified activity-silent decision-making dynamics also constitute a novel conceptual contribution. Our results show that activity-silent dynamics can organize decision formation: weakly expressed evidence accumulates before being rapidly converted into a categorical choice.

Learned STP organization. The STP assignments and parameters in Masse et al. (2019) are predefined, whereas our network learns heterogeneous STP strengths during task optimization. The resulting activity-silent decision dynamics therefore emerge through learning rather than being imposed as an operating regime.

A distinct circuit mechanism for decision formation. Previous models used STP to retain an earlier stimulus or an already formed choice (Deco et al., 2010; Martínez-García et al., 2011), amplify afferent evidence (Bourjaily & Miller, 2012), or drive transitions among discrete attractors (Ballintyn et al., 2019). In contrast, we link the evolving synaptic state during evidence accumulation to stability changes in a choice-selective E/I circuit and construct an independent theoretical model showing how these changes promote the transition to commitment. Thus, the contribution is the discovery and mechanistic analysis of an activity-silent evidence-to-choice transition, rather than the application of an established working-memory mechanism to a new task.

2. Direct evidence linking STP to evidence accumulation (W2)

Thank you for this comment. We directly tested STP dynamics in ten trained networks. In the fixed-weight interventions, we kept the trained input/recurrent/output weights fixed, but set effective synaptic efficacy q to one or retained only facilitation or depression, and then reevaluated decision performance. Separately, we decoded cumulative left-minus-right sensory input from effective synaptic efficacy q, neuronal activity r, or their combination [q,r]:

TestResult
Effective synaptic efficacy set to oneFinal-choice accuracy decreased by 12.21 points overall (95% CI for the paired change, -20.55 to -3.87; p=0.0059) and by 19.40 points on low-coherence trials.
Facilitation only (depression removed)Final-choice accuracy decreased by 36.55 points in the raw ten-network summary; four networks had unstable outputs.
Depression only (facilitation removed)Final-choice accuracy decreased by 11.76 points across ten networks.
Decode cumulative left-minus-right sensory inputHeld-out R²: neuronal activity r, 0.8930±0.0207; synaptic gain q, 0.9492±0.0136; joint [q,r], 0.9613±0.0098 (mean ± SD across ten networks).

The interventions show that changing STP-mediated transmission substantially impairs decisions. Specifically, facilitation-only caused output instability in four networks. Complementarily, q decoded cumulative evidence better than r in all ten networks and reduced RMSE by 31.2±7.2%. Together, these results directly link STP dynamics to evidence accumulation in the trained circuit. We will add the protocols and per-network results to the revised manuscript.

AuthorsOpenReview ↗

Authors reply: Reviewer ueQF (part 2/3)

3. Frozen-network stability analysis (W3)

Thanks for your comments. We agree that the frozen-network analysis is not equivalent to solving the full dynamics with evolving STP. However, this analysis is not intended to provide an exact trajectory of the dynamic network, but to offer an intuitive, quasi-static account of how the current synaptic state reshapes the network’s stability landscape. This approximation is motivated by the separation of timescales: neural activity evolves with τE=20 ms and τI=5 ms, whereas the dominant STP facilitation variable evolves with τu = 1500 ms (τx = 50 ms for depression). Neural activity can therefore approximately relax to a steady state conditional on the more slowly changing synaptic state. Treating the slow variable as a bifurcation parameter is a standard slow–fast analysis used in neural networks with dynamic synapses (Barak and Tsodyks, 2007; Aguilar et al., 2017; Taher et al., 2022). Importantly, our population-dynamics results are obtained from the full network with continuously evolving STP; the frozen analysis is used only to interpret how these synaptic changes can promote the transition toward a choice state. We will revise the main text to state this relationship more explicitly.

4. Stability transition in the trained RNN (Q2)

Thank you for this important question. The trained RNN is not mathematically equivalent to the four-population theoretical model, and we do not intend the latter to be a reduced or parameter-matched model of the trained RNN. Direct bifurcation analysis of the original N-dimensional RNN is difficult because its effective connectivity changes continuously with the STP variables, producing a high-dimensional, time-dependent dynamical system. This difficulty is precisely why we introduce a separate theoretical model. The model is inspired by two observations from the trained RNN: task optimization recruits STP, and pruning reveals a choice-selective E/I organization. The theoretical model isolates these ingredients and shows, in principle, how slowly evolving synaptic efficacy can reshape the stability of a competitive decision circuit and promote a transition to a choice state. We will clarify the models' relationship in the revised manuscript.

We tested the reviewer's proposed temporal prediction directly: can changes in effective synaptic gain anticipate the neural choice transition? In each of the ten trained networks, we defined the transition as the time of maximum rise in the choice-output margin. A decoder used only the preceding 100-ms change in effective synaptic efficacy, Δ q(t)=q(t)-q(t-100 ms), to predict whether this transition would occur in the next 100 ms. Held-out AUC was 0.652±0.020 across networks (range 0.621–0.681; 10/10 above chance; p=0.002). Average precision was 0.268±0.027, exceeding class prevalence by 0.106±0.027 in all ten networks (p=0.002). Thus, changes in synaptic efficacy contain prospective information about the timing of the neural transition.

AuthorsOpenReview ↗

Authors reply: Reviewer ueQF (part 3/3)

5. Computational advantage and controlled activity cost (W4, Q3)

Thanks for pointing this out.

For the computational advantage: Although Static achieved slightly higher final-choice accuracy on the validation set in the original comparison, this endpoint argmax metric is not the same as the masked trajectory loss used to select checkpoints, nor does it account for timeouts or response timing. It also depends on the STP setting; with U×2 (doubling baseline utilization U), validation accuracy reached 95.14±1.13%, close to the Static value. We therefore also evaluated held-out behavior and decision dynamics on the same 2,000-trial protocol (mean ± SD):

EvaluationSTPStatic
final-choice accuracy95.05±1.17%79.92±6.78%
timeout fraction5.46±2.26%13.96±3.26%
timeout fraction, low coherence (1 and 2)13.55±5.73%30.26±5.62%
Correct threshold crossing within the trial, all trials90.25±1.75%83.44±3.14%

On this held-out testing task stream, RNNs with STP had higher final-choice accuracy, fewer timeouts in all ten network pairs, and more correct threshold crossings, especially on weak-evidence trials. This identifies an online benefit for timely, well-separated commitment that final-output accuracy misses. Training also strengthened STP, and the higher step-like fractions indicate that learned STP is an inductive bias toward the activity-silent dynamics. The fixed-weight ablations in our response to W2 above further show that disrupting effective synaptic gain impairs decision performance. Complementary decoding shows that synaptic state carries a more accessible accumulated-evidence trace than neuronal activity. Together, these results support a role for STP in evidence processing and reliable within-trial commitment.

For the energy comparison: Action potentials and postsynaptic signaling account for major components of neural energy use, so integrated firing rate is a biologically motivated activity proxy (Attwell and Laughlin, 2001; Howarth et al., 2012). Moreover, metabolic costs are difficult to quantify. The variables u and x represent release probability and available synaptic resources, respectively. Neurotransmitter release and vesicle recycling occur at the biological synapses represented by both Static and STP conditions and incur metabolic costs in both (Tsodyks and Wu, 2013).

To perform the energy comparison under more closely matched conditions, we further assessed accuracy, reaction time, and expected decision-period spikes together. In the four-population model, we matched both accuracy and reaction time at three behavioral tiers. Each entry below is STP / Static:

Target-accuracy tierObserved accuracyMean reaction time (ms)Expected spikes per trial; reduction with STP (95% CI)
85%84.24% / 83.80%264.9 / 266.64.723 / 6.567; 28.08% (27.72%–28.46%)
90%89.01% / 88.70%410.2 / 414.87.675 / 10.269; 25.26% (24.84%–25.70%)
95%94.45% / 94.83%747.4 / 748.915.160 / 18.472; 17.93% (17.69%–18.17%)

Across the three tiers, accuracy differed by at most 0.44 percentage points and mean reaction time by only 0.19%–1.10%; STP nevertheless required fewer expected decision-period spikes at every tier. Thus, across closely matched behavioral operating points, STP achieved comparable accuracy and decision speed with consistently lower spiking activity. We will add the controlled sweep to the revised manuscript.

Thank you again for your valuable suggestions, which helped us sharpen both the mechanistic and activity-cost claims. We hope we have adequately addressed your concerns, and please feel free to let us know if anything remains unclear. We sincerely look forward to your feedback.

Reviewer_YAZXOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the use of short term plasticity (STP) in Dalean recurrent neural network models as a means to emulate activity-silent working memory function when the network performs a decision making task (reaction -time motion-discrimination task).

The authors equip each neuron with short term plasticity variables (the rate-based Tsodyks-Makram model, that considers both short term facilitation and depression), and optimise how strongly the short term plasticity dynamics influence the outgoing connections of each neuron by training the variables aja_j, together with the recurrent connection strengths. They train the networks on behavioural variables of a decision making task, and compare the trained STP-equipped networks to conventional ones that have only trainable weights.

They find that the STP network develops more step-like activity than the conventional RNN. (The authors employ a detailed fitting framework to determine whether single neuron trajectories are steps-like or not, but I did not get into detail there).

Then the authors construct a minimal model with four populations (2 E+2 I) to study how activity dependent changes in the effective synaptic weights affect the dynamical landscape of this reduced model. They show that by increasing the synaptic gain between the population they change the fixed-points in the system and allows for transitions between monostable-bistable, and trustable dynamics. To construct the minimal model the authors identify the choice-selective neurons, remove connection s between different choice-selective neurons and measure the performance decrease, and subsequently pinpoint the neurons with large learned STP variables aa and characterise how important they are for decision making. However, as I understand the minimal model has only coarse qualitative characteristics of the sub-network of the RNN they identify as "important" for the considered computation.

Critical strengths
  • I find the proposed mechanism of evidence accumulation through synaptic dynamics very intresting, and this is the strongest point of the paper.

  • I appreciate that the authors did not stop into the training the recurrent network s but they subsequently performed further analysis and provided additional theoretical results.

  • I appreciate the functional ablation experiments where the authors removed the choice selective sub-populations

  • I liked the proposal of the minimal model to demonstrate the bifurcations that potentially happen in the system as the synaptic dynamics evolve (however see my question below)

Critical weaknesses
  • The authors do not provide any sort of sensitivity analysis to probe whether their findings survive when the pre-selected short term plasticity time constants and other variables change within physiological ranges.

  • As I understand, the ablations the authors provide show that connections/neurons with stronger STP variables are more choice selective and by silencing these neurons the network performance deteriorates, but this doesn't necessarily imply that the evidence is accumulated in the synaptic variables uu and xx. I would expect a more thorough analysis that would probe whether indeed one of those variables stronger "correlates" with evidence.

  • The theoretical exposition and the trained RNN use different STP parameters, but the authors do not clearly demonstrate that these are equivalent.

Questions for authors
  1. How robust are the results to the selection of the specific values of the facilitation and depression time constants and to the baseline utilisation variable UU, or initial values of the strength aja_j? Have you performed some sort of robustness check or can you provide evidence that within physiologically meaningful variations of these parameters the results do not change qualitatively?

  2. Can you provide some visualisation s of the trained weights together with the associated trained STP strengths aa, and can you analyse how or whether synaptic strength correlates with STP strength?

  3. Can you demonstrate that the versions of the model you explore numerically and theoretical are indeed equivalent, either by performing training on the version of the network you analyse theoretically or the reverse. I appreciate the existence of both theoretical exposition and nummerical results but the two should consider the same model or demonstrate that the two models are in some sort equivalent.

  4. Can you perform ablations wrt to facilitation and depression, to probe whether one form of short term plasticity is more necessary to explain the step-like activity?

  5. If you would compare a decoder trained on the synaptic variables (either both uu and xx or each one of them), and would compare with a decoder trained on neuronal activity would the first set of decoders show a better performance than the neuronal activity based one?

  6. I am not sure that I understood how you construct the minimal model, i.e. select the parameters etc from the trained RNN? Are they supposed to be connected with some sort of reduction or not?

8 comments
AuthorsOpenReview ↗

Authors reply: Reviewer YAZX (part 1/3)

Thank you for your careful and constructive review. We appreciate your positive assessment of the synaptic evidence-accumulation mechanism, functional ablations, and theoretical analysis. Below, we present additional analyses and clarify the precise relationship between the trained network and the theoretical model.

1. Sensitivity to STP parameters and initialization (W1, Q1)

Thank you for pointing this out. The manuscript already provides evidence that the result is robust to STP-strength initialization. In Figure 2c, the network-mean STP strength increased from 0.104 to 0.587, from 0.501 to 0.697, and from 0.898 to 0.938 for different initial means. The increase was positive in all 30 networks. At every initialization, the activity-silent fraction exceeded that of the same-seed Static RNNs without STP across 10 seeds (exact paired p=0.00195):

Initial mean STP strengthIncrease in network-mean STP strengthSTP activity-silent fractionMatched Static fractionPaired activity-silent increase
μ=0.1+0.484±0.0300.586±0.0150.401±0.016+0.184±0.022
μ=0.5+0.196±0.0230.625±0.0120.409±0.019+0.215±0.024
μ=0.9+0.040±0.0040.568±0.0140.409±0.016+0.159±0.023

To test the sensitivity of other hyperparameters in STP, we retrained ten seeds for each 0.5-fold or 2-fold change in three quantities (60 networks in total). Baseline utilization U sets the resting fraction of synaptic resources used by activity; τᵤ sets how quickly facilitation decays toward baseline; and τₓ sets how quickly depleted resources recover. In all trained networks, the main results remained qualitatively unchanged. Specifically, network-mean STP strength increased in all 60 networks, and the activity-silent fraction exceeded the same-seed Static RNN without STP under every parameter condition across 10 seeds (exact paired p=0.00195):

Parameter conditionIncrease in network-mean STP strengthPaired activity-silent increaseValidation accuracy (mean ± SEM)
U×0.5+0.422±0.031+0.314±0.02089.04±0.32%
U×2+0.416±0.016+0.084±0.01795.14±0.36%
τᵤ×0.5+0.517±0.016+0.210±0.02792.72±0.62%
τᵤ×2+0.485±0.033+0.219±0.01691.93±0.59%
τₓ×0.5+0.494±0.026+0.181±0.02492.62±0.65%
τₓ×2+0.480±0.030+0.190±0.02392.86±0.58%

Together, these results show that the conclusion does not depend on a single hand-picked STP setting. We will add the complete condition-wise results to the revised manuscript.

2. Contribution of synaptic variables and component ablations (W2, Q4)

Thanks for raising this important question.

We intervened on STP dynamics while preserving the inputs, recurrent and readout weights. Specifically, we set effective synaptic efficacy to one (q=1), retained facilitation alone by setting x=1, or retained depression alone by fixing u=U. We measured final-choice accuracy in ten trained networks. Values are mean ± SEM across networks; the change is paired to intact STP:

Same-weight conditionFinal-choice accuracy (paired change)
Intact STP94.34±0.74%
Effective synaptic efficacy set to one82.13±3.35% (-12.21±3.69 points)
Facilitation only (depression removed)57.79±4.29% (-36.55±4.14 points)
Depression only (facilitation removed)82.58±2.46% (-11.76±2.77 points)

All three interventions reduced mean accuracy across the ten networks. Facilitation-only caused the largest decline and produced extreme output amplitudes in four networks. These fixed-weight results show that STP-mediated transmission contributes to the trained decision computation. Neither component alone reproduced intact performance.

To examine the neuronal transition directly, we also fit a sigmoid to choice-aligned high-STP neurons during the first 600 ms after stimulus onset in each of the ten networks. For each trial, we compared the mean signed midpoint slope of cells meeting the fit criteria under both intact STP and the component ablation; the table reports the ablation-minus-intact difference, averaged across network-level medians (mean ± SEM, firing-rate units/s):

Same-weight conditionSigmoid slopePaired slope changeNetworks included in slope fit
Intact STP6.043±0.572Reference10/10
Effective synaptic efficacy set to one6.317±0.243-0.299±0.25310/10
Facilitation only (depression removed)2.306±1.449-1.069±0.7986/10
Depression only (facilitation removed)6.736±0.134-0.881±0.37510/10

The discussion of these slope results continues in part 2/3 below.

AuthorsOpenReview ↗

Authors reply: Reviewer YAZX (part 2/3)

2. Contribution of synaptic variables and component ablations (W2, Q4) (continued)

Continuing Q2 from part 1/3: Facilitation-only severely disrupted the trained dynamics: 4/10 networks developed extreme readout amplitudes (maximum absolute output ranged from 11,573 to 1.71 × 10³⁰), so their slopes are excluded. The six networks included in the slope fit still had mixed slope changes. Setting efficacy to one produced a small paired slope change (8/10 networks decreased). Depression-only also remained numerically stable and decreased the paired slope in 9/10 networks. Together, these interventions indicate that both facilitation and depression shape the neuronal transitions associated with step-like activity. In particular, facilitation without counterbalancing depression can greatly amplify the network response and disrupt its dynamics.

Thank you again for this valuable comment. We will add the per-network accuracy and slope-fit results to the revised manuscript.

3. Correspondence between the trained RNN and theoretical model (W3, Q3, Q6)

We apologize for not making the relationship between the two models clearer. The theoretical model is constructed for modelling the mechanism of activty-silent decision-making under a clearer frame, which is inspired by the RNN analysis, but not a reduction or parameter-matched equivalent of the trained RNN. Thus the direct answer to your question Q6. “Are they supposed to be connected with some sort of reduction or not?” is No.

Specifically, the RNN provides mechanistic and architectural inspiration:

(1) Task optimization recruits STP, which implies that STP serves as an important inductive bias for generating activity-silent decision-making dynamics;

(2) Pruning reveals a choice-selective opponent E/I four-population circuit in which excitatory populations reinforce their own choices while inhibitory populations suppress the competitor.

These findings motivate an independent four-population theoretical model that asks whether STP is sufficient to transform the dynamics of a competitive decision circuit. Combining these two inspirations, together with classical competitive decision circuit like in Wang (2002) and Wong and Wang (2006), with the E/I organization also informed by Kuan et al. (2024), we build the theoretical model. Its connection to the trained RNN is mechanistic and architectural, not a fitted reduction or parameter-matched equivalence. We will make this distinction explicit in the revised manuscript.

Construction and parameter robustness of the minimal model (Q6)

Without STP, the four-population model is a classical competitive decision circuit based on Wang (2002) and the population-rate formulation of Wong and Wang (2006). Its recurrent competition and attractor dynamics are well established. The baseline parameters are chosen according to these classical works, and they were not fitted to the trained RNN.

We then add STP to test whether evolving synaptic efficacy can generate a prolonged low-activity period followed by rapid choice commitment. We first performed a bifurcation analysis and identified three regimes as synaptic efficacy increased: monostability with only a low-activity state, tristability with one low-activity and two choice states, and bistability with only the two choice states. The tristable regime allows the network to remain at low activity during evidence accumulation, whereas entering the bistable regime destabilizes the low-activity state and forces the network toward a choice state. We therefore hypothesized that activity-silent decision formation corresponds to a transition from tristability to bistability: the circuit can initially remain in the low-activity state, but increasing synaptic efficacy eventually destabilizes this state and allows the activity to switch rapidly to one of the two choice states. Based on this phase diagram, we selected parameters for which the evolving synaptic efficacy crosses the tristable-to-bistable boundary during a trial, and then verified the predicted state transition in simulations of the full dynamic STP model.

AuthorsOpenReview ↗

Authors reply: Reviewer YAZX (part 3/3)

4. Relationship between trained connectivity and STP strength (Q2)

We thank the reviewer for suggesting a direct comparison of trained weights and STP strengths. We analyzed associations with sensory-input, recurrent, and readout weights in all ten independently trained networks (seeds 201–210). The table reports the mean of ten within-network Spearman correlations, and statistical tests use networks as independent observations.

Relationship with trained STP strengthMean Spearman ρConsistent networksTwo-sided p
Magnitude of left–right sensory-input selectivity+0.57410/10 positive0.00195
Outgoing recurrent-weight L₂ norm-0.14910/10 negative0.00195
Magnitude of choice-output selectivity+0.0647/10 positive0.160

Thus, stronger STP is associated with more selective sensory evidence input. Its association with outgoing recurrent-weight magnitude is consistently negative, whereas the association with output selectivity is weak and not reliable.

Since we hypothesized that evidence accumulates in neurons with stronger STP, this experiment provides more evidence. Thank you for helping us improve this work. We will add the visualisation and analysis results to the revised manuscript.

5. Synaptic-state versus neuronal-state decoding (W2, Q5)

We thank the reviewer for asking whether synaptic state carries accumulated evidence more accessibly than neuronal activity. To test this directly, we decoded the realized cumulative difference between the two sensory inputs, including input noise, from utilization u, available resources x, effective synaptic efficacy q (computed from u and x), firing rates r, and their concatenation [q,r]. In each of ten trained networks, we split complete trials into 1,600 training and 400 held-out trials. For each representation, one linear decoder was fitted using all 75 sampled times of the training trials and then applied to the held-out trials. The trained network, rather than a trial or time point, is the unit of comparison. Values are mean ± SD across networks:

DecoderHeld-out cumulative-evidence R²RMSE (input units × s)
Neuronal activity r0.8930±0.02070.0753±0.0072
Effective synaptic efficacy q0.9492±0.01360.0517±0.0065
Utilization u0.9645±0.00670.0433±0.0038
Available resources x0.9532±0.00800.0498±0.0040
Joint [u, x]0.9698±0.00550.0400±0.0034
Joint [q,r]0.9613±0.00980.0452±0.0054

Both synaptic variables individually outperformed neuronal activity in all 10 networks: the mean paired Δ R² was +0.0716 for u and +0.0603 for x (two-sided Wilcoxon p=0.00195 for each). Utilization u also exceeded x in all 10 networks (Δ R²=+0.0113, p=0.00195), while combining them further improved performance to R²=0.9698 in all 10 networks. Thus, the accumulated evidence is more linearly accessible from either synaptic variable than from neuronal activity, with u providing the stronger individual readout and [u, x] providing the strongest readout overall. We will add the decoder protocol and per-network results to the revised manuscript.

Thank you again for your constructive comments, which have substantially improved our work. We hope we have addressed your concerns. Please let us know if anything remains unclear. We look forward to your feedback.

Comments on the paper

5 comments
AuthorsOpenReview ↗

To All Reviewers

We thank all reviewers for their careful and insightful feedback. Your comments prompted additional analyses that strengthen our central conclusion: STP supports evidence-guided decisions with activity-silent transitions. These analyses have made our conclusions more rigorous and comprehensive and deepened our understanding of the underlying mechanism. We will incorporate the new results and clarifications into the revised manuscript as soon as possible.

Two concerns recur across the reviews: whether the trained RNNs provide direct evidence linking STP to evidence accumulation, and how they relate to the theoretical model. We addressed the first through fixed-weight interventions and decoder comparisons. Disrupting effective synaptic efficacy impaired decisions, especially on weak-evidence trials, and synaptic variables provided a more accessible linear readout of cumulative sensory input than neuronal activity. Regarding the second, the four-population theoretical model is an independent mechanistic model inspired by the trained RNN, rather than a fitted reduction. It draws on two observations: learning strengthens STP, and pruning reveals a choice-selective E/I organization. The model tests how changing synaptic efficacy could drive a stability transition and trigger rapid commitment to a choice. Additional task tests, comparisons with Static networks, and behavior-matched activity-cost analyses provide further support for our conclusions. The individual replies present the details. We would like to sincerely thank the reviewers once again for their constructive comments, which have substantially improved our work.

AuthorsOpenReview ↗

Manuscript Revised

We have now uploaded the revised manuscript. The main additions and revisions are highlighted in red. For ease of reference, the principal updates are:

  • Direct evidence in trained RNNs: fixed-weight STP interventions and evidence-decoding comparisons in Sections 3 and 4.1, Figures 2g and 3a, and Appendices A.7 and A.8. Appendix A.7 also includes facilitation/depression ablations.
  • Robustness and connectivity: STP parameter and initialization tests in Appendix A.6, and the relationship between trained connectivity and STP strength in Section 4.2 and Appendix A.11.
  • Mechanistic clarification: the relationship between the trained RNNs and the independent theoretical model in the Introduction, Section 5, and Discussion, with the interpretation of frozen-synapse analysis clarified in Section 5.2 and Appendices B.2 and B.3.
  • Controlled activity-cost comparison: expected spike counts at comparable accuracy and reaction time in Section 5.2, Figure 5e, and Appendix B.5.

We thank the reviewers again for their constructive suggestions, which helped improve the manuscript. We hope these additions addressed your concerns; if so, we would be grateful if you would consider raising your score. Looking forward to your reply.

Thank you very much for your positive feedback and helpful suggestion. We are glad that the additional experiments and explanations addressed your concerns. We will increase the font sizes in all figures to improve readability in the revised manuscript.

A small note for clarity: your comment appears under our response to Reviewer ueQF. We wanted to flag this to avoid any confusion about which review the feedback refers to.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.