acceptodds
Under review as a conference paper at ICLR 2027

Dendritic Computation Shapes How Networks Learn to Use Context

16 views

Abstract

Flexible behavior requires learning how to use sensory and contextual information. A recent in vivo study found that activating dendrite-inhibiting interneurons impaired rule relearning while sparing established behavior. How dendritic input organization shapes learned sensorimotor mappings remains unclear. Here we develop a trainable framework of recurrent circuits built from biophysically detailed neurons to manipulate input location and dendritic conductance during learning and execution. In a controlled task, sensory input encodes both direction and rule, while a separate context stream repeats the rule. Apical and basal routing yield similar performance with all inputs present, but apical routing produces less dependence on context after learning. Exchanging sensory weights after training and sharing them during training show that learned sensory parameters contribute to both the expression and formation of this difference. Blocking the apical calcium conductance during training partially narrows the post-withdrawal performance gap even after conductance restoration, whereas acute blockade has much smaller effects on performance without context. Within apical-context circuits, exchanging sensory weights between these training histories also transfers this persistent effect, linking conductance conditions during learning to functionally consequential sensory parameters. A simulated robotic pushing task further shows that context availability during learning affects subsequent closed-loop control under common execution conditions. These findings identify dendritic computation as a source of learning bias, linking cellular input organization and biophysics to learned sensorimotor mappings, with a proof of concept in artificial control.

Primary area:
applications to neuroscience & cognitive science
Keywords:
dendritic computation, sensorimotor learning, context dependence, biophysically detailed neural networks, biophysical neuron modeling, 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_4x4COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

In this paper the author investigate the impact of input segregation on the resulting computation learned by a small network of morphologically realistic neurons trained to perform sensorimotor tasks. The author start from a recent experimental finding in mice where induced silencing of apical dendrites resulted in impaired relearning, while established behaviour was left mostly unaffected. To computationally study the impact of dendritic organisation on the kind of solution learned by a network, the authors trained a network a biologically realistic neurons on a rule-following task where sensory information was routed to the basal dendrites while additional context (informationally redundant with the sensory signal) was routed either to the basal or apical dendrites during training. At test time the context signal was withdrawn, and the network with input segregation (i.e. context in the apical compartment) showed significant reduced performance degradation. A subsequent interventional experiment in which the input sensory connections were swapped between the two configurations, illustrated how such sensory connections were indeed largely responsible for the different performance losses, with a sensory weight-matching training configuration further validating the point as the resulting configuration almost closed the gap between the two configurations. Lastly, the author proposed a proof-of-concept demonstration in a closed-loop task where input segregation resulted in better performances.

Critical strengths

The work opening motivation (a recent experimental finding on the impact of apical dendrites deactivation on mice behaviour) is intriguing and the subsequent experimental design is well crafted, with the two explored configurations (basal and apical context input) evaluated on subsequent test-time withdrawal of the context signal. The subsequent interventional experiments help locate the source of the observed difference in the sensory connections that training sculpts depending on context segregation. The final more realistic task is a nice addition towards more realistic closed-loop scenario.

Overall, the framing of dendritic organisation as a form of inductive learning bias is a valuable (and novel) contribution, which would be interesting for the field. Finally, the paper is generally well written with polished figures and a detailed appendix.

Critical weaknesses

One important limitation of the work concerns the various limits inherited from the complexity of simulating a network of morphologically-extended biologically realistic neurons. One obvious limit concerns the scale of the explored networks (4040 neurons) and the simplicity of the main task (although the authors do include the more realistic CoffePush task, which in turn however did result in significantly reduced differences; Fig.5C). A second limitation concerns the difficulty in interpreting the result due to the many degrees of freedom of the considered neural models: Fig. 2B shows a dependency between the apical-basal gap on one of the model parameters (sCs_C), and it thus remains unclear how much (or whether) the observed gap is robust to changes in other model parameters that have a range of biologically plausible values (e.g. does an increase in overall firing rates alters the picture?)

Questions for authors
  1. Does the proposed principle of dendritic organisation as a learning bias extend to simplified multi-compartment neurons (e.g. Poirazi & Papoutsi Nat. Rev. Neuro. 2020; Capone et al. PNAS 2023; Graham et al. Neural Comp. 2025)? Or is there some important aspect that is lost by abstracting away the morphology?
  2. In the main task, the context is redundant so that even without context the network learns the task equally well (Fig. 2A). What happens if the context is not redundant? The CoffeePush task (Fig. 5B) seems to be hinting to the fact that input segregation generally helps performances, while the gap itself becomes more comparable between the two.
  3. How dependent are the findings (observed learning bias) on the choice of the learning algorithm (e.g. e-prop Bellec et al. Nat. Comms. 2020)? Or do these differences persist with longer training?
5 comments
AuthorsOpenReview ↗

Response to Weakness

Thanks for your positive evaluation and for highlighting our framing of dendritic input organization as a learning inductive bias. We agree that network scale, task complexity, and parameter robustness are important limitations to address. In the revised manuscript, we have added comparisons across neuronal abstraction levels, analyses of different firing-rate regimes and extended training, and a second robotic manipulation task. All new additions are highlighted in blue.

Major Concern: Network Scale, Task Complexity, and Parameter Robustness

We thank the reviewer for recognizing the computational cost and simulation complexity associated with networks of morphologically detailed neurons. Because simulating multicompartment cable equations, active ion channels, and trial-level gradient replay is computationally expensive, our main circuit is currently limited to 32 excitatory and 8 inhibitory neurons, and we leave larger networks for future work.

Meanwhile, to address the reviewer's specific concerns, we have combined robustness analyses already present in the original manuscript with newly added experiments to broaden the model, parameter, and task coverage along several dimensions:

1. Parameter robustness: context learning rates and weight-update conditions (Appendix B, Fig. S5, already included in the original manuscript).

The original manuscript already examined context-weight learning rates set to 0.5, 1, and 2 times the baseline value, as well as a control condition in which context weights were frozen while all other parameters continued to be updated. The basal-minus-apical withdrawal-cost gaps were 0.04830, 0.05451, and 0.07600 at the three learning rates, respectively, and remained at 0.03741 when context weights were frozen. Although the magnitude of the difference varied with the learning configuration, the lower context dependence of apical routing was preserved across all these conditions. These existing results provide a basis for evaluating the influence of learning-parameter variations on our main finding.

2. Robustness across different levels of network activity (new Appendix H.1, Fig. S18).

To address the reviewer's specific question about firing rates, we varied the sensory input event rate to 0.75, 1, and 1.5 times its baseline value. The resulting mean population firing rates with context present after training were approximately 6, 11, and 24 Hz, respectively, covering substantially different levels of network activity. Under these three conditions, the mean basal-minus-apical withdrawal-cost gaps were 0.1163, 0.0545, and 0.0402, respectively. All three paired initializations retained the same direction of the routing difference. Increasing sensory drive reduced the magnitude of the difference but did not reverse its direction.

3. Robustness to longer training (new Appendix H.2, Fig. S19).

We extended training of the main model from 500 to 1,000 epochs, resuming from the epoch-500 checkpoints with their network weights and optimizer states restored. The mean basal-minus-apical withdrawal-cost gaps at epochs 500, 750, and 1,000 were 0.0545, 0.0541, and 0.0551. All three network pairs maintained a lower withdrawal cost for apical routing. Thus, the learning bias observed in our main experiments is not restricted to a particular training endpoint.

4. Extension to a second embodied manipulation task (new Appendix I, Fig. S20, Table S12).

To address the relative simplicity of the main task, we added a second MetaWorld manipulation task, Hand-insert, which involves three sequential stages: approach and grasp, transport and alignment, and lowering and insertion. This task involves action requirements and closed-loop interactions that differ from those in CoffeePush. In Hand-insert, the mean context-present success rate was 41.64% for apical routing and 37.67% for basal routing. After context withdrawal, the success rates decreased by 2.06 and 5.34 percentage points, respectively. Across all three seeds, basal routing consistently exhibited a higher withdrawal cost than apical routing. This extends our closed-loop control results from a single CoffeePush task to a second manipulation task.

AuthorsOpenReview ↗

Response to Question 1: Simplified multicompartment models and what morphological abstraction loses

The review by Poirazi and Papoutsi (2020) systematically discusses dendritic computation models at different levels of abstraction. Following the abstraction hierarchy, we tested an abstract gain network (Wybo et al., 2023), a three-compartment spiking model (Capone et al., 2023), and a reduced 20-compartment L5 pyramidal cell model (Bahl et al., 2012; Graham et al., 2025) on the same task (new Appendix F, Figs. S14–S15, Table S11). Simplified models fail without active dendrites, whereas a reduced biophysical model with active dendrites works:

Gain and three-compartment models give the opposite pattern. Without the cue, modulatory/apical routing performs worse than additive/basal routing in both the gain model (withdrawal-cost gap = -0.2315) and the three-compartment model (withdrawal-cost gap = -0.1387 under BPTT; -0.1365 under e-prop). Input segregation or gain modulation alone is therefore not enough.

An active 20-compartment biophysical model reproduces the apical advantage. Bahl et al. (2012) compressed a reconstructed L5 pyramidal neuron into seven anatomical regions (20 compartments) optimized against experimental voltage responses, preserving dendritic Ca2+ spikes and soma–dendrite coupling. Retaining active channels and AMPA/NMDA/GABA synapses reproduces our main result: cue removal raises apical NMSE only from 0.0468 to 0.0500, versus 0.0335 to 0.0833 for basal routing (withdrawal-cost gap = +0.0466).

Removing active dendritic channels erases most of the advantage. Making dendrites passive in the 20-compartment model—while keeping compartment separation, axial coupling, and NMDA synapses—cuts the withdrawal-cost gap to 0.0210, and removing only apical tuft Ca2+ channels cuts it to 0.0092 (Appendix F.3, Fig. S15B,D). Abstracting away active dendritic dynamics thus loses the core mechanism; how fine branch topology in the full Larkum et al. (2009) morphology further shapes learning remains an open question for future work.

Taken together, our results support a more specific conclusion: the learning bias induced by dendritic input organization depends on the combination of spatial input organization and local active dendritic dynamics. Abstract gain models and three-compartment models can retain computational features such as input modulation or spatial segregation, but these features alone are insufficient to determine the context dependence acquired through learning.

AuthorsOpenReview ↗

Response to Question 2: Nonredundant context and the interpretation of robotic control results

The informational relationship between sensory input and context determines what withdrawal measures. We distinguish fully redundant context, task-essential context, and the partially redundant phase cues used in the robotic tasks.

Fully redundant context isolates learned reliance on an additional pathway. In the main task, sensory input already encodes both direction and rule, and context repeats the rule. The task remains solvable after withdrawal, which reveals different learned solutions despite comparable full-input performance. Post-training sensory-weight exchange transfers the behavioral effect, while sharing sensory weights during training reduces it (Fig. 3). This connects the withdrawal phenotype to how the sensory mapping was acquired.

Task-essential context changes the meaning of the same intervention (new control in Appendix B, pp. 22–23). When the same sensory stimulus requires different actions under different rules, context supplies information needed to select the response. Auditory rule-switching tasks exemplify this structure (Zhang et al., 2026). We implemented this setting by disconnecting input from the 20 rule-encoding sensory channels, retaining direction input and the remaining training and evaluation settings.

Both routes learn the task with context present, but their errors rise to approximately 1.0 after withdrawal:

RoutingContext-present NMSEPost-withdrawal NMSENMSE increase
Apical0.059 ± 0.0021.040 ± 0.0020.981 ± 0.004
Basal0.056 ± 0.0111.022 ± 0.0020.966 ± 0.012

Here, withdrawal removes the rule information needed to determine the target. Similar full-input errors do not imply identical representations or learning dynamics. Under task-essential context, those differences would be better examined through learned responses, synaptic organization, or acquisition dynamics, because the post-withdrawal error is dominated by missing task information. The redundant-context design makes that learning question directly accessible through behavior.

The robotic tasks combine substantial sensory–context overlap with explicit stage information (Appendices D, I). CoffeePush and Hand-insert require continuous control of the geometric relationships among the end effector, object, and target. Their phase cues identify stages such as reach/contact/push or approach-and-grasp/transport-and-align/lower-and-insert. Linear classifiers recover phase labels from current sensory observations with cross-validated balanced accuracies of 97.7% and 97.33%, respectively, on the expert-demonstration data (Appendices D.2 and I.2). Stage information is therefore substantially accessible through sensory input, while the explicit cue also incorporates trajectory history.

Knowing the stage does not specify the complete action: within each stage, the controller must still map continuous geometry to movement. Context can consequently influence both current action selection and how sensory–action mappings are organized during learning. Consistent with this interpretation, apical routing shows higher mean cue-present success and smaller mean withdrawal costs in both tasks (means across three networks; withdrawal costs in percentage points, pp; Tables S7, S12):

TaskApical success with cue (%)Basal success with cue (%)Apical withdrawal cost (pp)Basal withdrawal cost (pp)
CoffeePush49.4045.841.373.46
Hand-insert41.6437.672.065.34

Hand-insert extends the comparison to different action requirements. Its withdrawal cost is smaller for apical routing in every seed. Apical post-withdrawal success also exceeds basal post-withdrawal success and the independently trained context-free reference in every seed; the corresponding means are 39.58%, 32.33%, and 37.09% (Table S12).

The CoffeePush continuations directly test a learning effect. At epoch 20, each source controller is duplicated with identical weights and Adam states, then continued to epoch 50 with context retained or withdrawn. With context present at evaluation, retaining it during continued training improves apical success by 9.42 percentage points, positive in all three sources, and reduces basal success by 2.13 percentage points, negative in all three sources (Fig. 5C, Table S6B). Thus, the contribution of stage context depends on its route during learning, even when execution-time input availability is held fixed.

These findings answer both aspects of your question: apical routing is associated with better overall performance in these partially redundant settings, and the paired continuations show that context availability during learning changes later control. Together, these tasks distinguish learned dependence, loss of necessary information, and closed-loop consequences.

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

The paper addresses the effects of dendritic inputs on contextual learning. Using a biophysical compartmental model of a small neural network with sensory and contextual inputs (rules), recurrent connections and readout weights, the goal is to better understand the effect of dendritic input on the task performance by neurons. Two learning strategies are compared,1-sensory information arrives at basal dendrite of the neuron and contextual information arrives at apical dendrites, and 2-both sensory and context information only arrive at the basal dendrite. In presence of both sensory and contextual inputs, both types of learning are successful and show no significant difference. When contextual input is removed, however, the learning that involves both basal and apical dendrites is more robust. The results are related to an observed phenomenon in the biological brain, where inhibiting apical dendrites impairs relearning of a new rule but maintains performance on an established rule.

Critical strengths
  1. Figures are telling and useful, and convey the main results of the paper.
  2. Exchanging sensory weights between the two trained networks is an insightful strategy to better understand the difference in processing between the network with and without input to apical dendrites.
  3. Blocking of calcium conductance in principle appears as a good intervention.
Critical weaknesses
  1. Writing is unclear, it does not emphasize important findings and does not explain clearly the results. I have found most of the paper unappealing to read, as the important and less important information is mixed together without much structure. The Methods beyond the first two subsections are particularly unclear and require substantial effort to extract any useful information. Variables such as WoW_o, bob_o, hh, yy, δz\delta_z, gSg_S are not explained. While some of them can be inferred, writing equations without explaining the variables is not a good practice and creates a substantial confusion.

  2. The mechanism of the effect is not explained. The setup does not seem too complex and in my opinion it should be possible to individuate the mechanism that would tell us why the network with and without input to apical dendrites learns differently.

  3. The significance of the contribution is only moderate, in particular due to the way this paper is written. The link with empirical observations in the biological brain are only loose.

Questions for authors
  1. Are "sensory weights" feedforward weights?

  2. Are WoW_o and bob_o the connection weights and the bias?

  3. How are sensory, recurrent and readout parameters learned jointly? This is unclear from Eq. 3 and in general, while it seems important. Line 207-208 mentions that Sensory, Recurrent and Context parameters are jointly learned, but the Eq. 3 only shows a variable W0W_0 multiplying the variable hh.

  4. In the literature, hh sometimes denotes the membrane potential, but in your Eq. (1), the membrane potential seems to be denoted by VV. This again I infer from the literature, because V(t)V(t) is ofter used for the membrane potential, but I do not see it spelled out in your Methods. So what does hh stand for? What is it relation to VkV_k?

  5. What is gˉS\bar{g}_S?

6 comments
AuthorsOpenReview ↗

Response to Weakness 1 — Part 1

Thanks for your thoughtful comments and for highlighting our figures, sensory-weight exchanges, and calcium-blockade experiments. In our main task, sensory input arriving on basal dendrites already carries both stimulus direction and rule, while a separate cue repeats the rule. Both circuits perform similarly while this extra cue is on (0.053 vs. 0.051 NMSE), yet when we remove it after training, the circuit that received the cue on apical dendrites makes far fewer errors than the one that received it on basal dendrites (0.145 vs. 0.197 NMSE; Fig. 2A). In short, where the extra cue arrives during training changes how well the sensory pathway learns to solve the task on its own. We rewrote Section 3 and added cross-model, gradient, and robotic experiments (marked in blue) to make both the math and the mechanism clear.

Weakness 1. How does the circuit learn, and what do the symbols mean?

We agree that the original presentation did not make the learning procedure sufficiently easy to follow. We have revised Section 3 to define the model variables explicitly and explain the sequence from sensory, contextual, and recurrent inputs to neuronal activity, action prediction, and joint parameter updates. Algorithm 1 now separates readout-gradient computation, conversion of spike-count credit into a voltage learning signal, synaptic gradient computation, and parameter updates. Appendix A.6 provides the corresponding equations and clarifies which parameters are learned and which remain fixed. We also distinguish synaptic gradients, denoted by G\mathcal G, from conductances. These revisions address the methodological questions below.

Q1. Are sensory weights feedforward weights?

Yes. Sensory weights SS connect external sensory spike channels to basal synapses on the circuit neurons. These are feedforward input connections. The circuit also contains recurrent weights RR, connecting neurons within the network, and context weights CC, connecting the external rule-cue channels to apical or basal dendrites of excitatory neurons. Sensory input encodes both stimulus direction and rule; the additional context stream repeats the rule cue (Section 3; Appendix A.1).

Q2. Are WOW_O and bOb_O the connection weights and bias?

WOW_O and bOb_O are the readout weight matrix and bias, respectively. They map the 40-neuron somatic spike-count vector hh to the two-dimensional action through y^=tanh⁡(WOh+bO)\hat y=\tanh(W_Oh+b_O). The sensory, context and recurrent synaptic weights are denoted separately by S,C,RS,C,R (Section 3, Eq. 3).

Our responses to Questions 3–5 are provided in the next comment.

AuthorsOpenReview ↗

Response to Weakness 1 — Part 2

Q3. How are sensory, recurrent, context and readout parameters learned jointly?

All parameter groups are optimized from the same task loss, although their gradients are computed differently.

First, in Eq. (3), hh is the vector of somatic spike counts produced by the network during a trial. Although the equation explicitly displays only the readout operation, hh itself depends on the sensory, context, and recurrent weights S,C,RS,C,R, which determine the circuit’s activity. Thus, these synaptic parameters influence the prediction through hh, while WO,bOW_O,b_O map hh to the predicted action.

Second, the task error provides a common learning signal. We compute the loss between the predicted and target actions. Standard differentiation through the readout gives its parameter gradients and assigns an error signal to each neuron's spike count. Because spike counts are discontinuous, we use the voltage-dependent surrogate described in Appendix A.6 to convert this error signal into a time-resolved somatic voltage learning signal:

=\mathbf 1_{n\in\mathcal W} \delta_{h,i}\\,s_i(V^{\mathrm{soma}}_{i,n}), \qquad \delta_h=W_O^\top[(\hat y-y)\odot(1-\hat y^2)].$$ Here, $\eta_{i,n}$ is the voltage learning signal for neuron $i$ at time step $n$; $\delta_{h,i}=\partial\mathcal L/\partial h_i$ is the loss sensitivity to its spike count; and $s_i$ is the voltage-dependent surrogate derivative specified in Appendix A.6. The voltage $V^{\mathrm{soma}}_{i,n}$ is recorded at the soma, $\mathcal W$ contains the time steps in the 50–500 ms readout window, and the indicator $\mathbf 1_{n\in\mathcal W}$ sets the signal to zero outside that window. The symbol $\odot$ denotes elementwise multiplication. Third, gradient replay computes the somatic voltage sensitivity to each synaptic weight. For each sensory, context, or recurrent synapse, it computes this sensitivity by replaying weight-specific gradient currents through the biophysical compartmental dynamics. Combining the resulting sensitivity with the corresponding voltage learning signal gives the task gradient supplied to the optimizer: $$\mathcal G_{w_e} =\sum_{n\in\mathcal W} \eta_{i(e),n}\\,\widehat V_{i(e),n\mid e}.$$ Here, $w_e$ is the weight of synapse $e$, $i(e)$ is its postsynaptic neuron, and $\widehat V_{i(e),n\mid e}$ is the somatic voltage sensitivity to $w_e$ obtained by gradient replay. Following He et al. (2026b), synaptic credit assignment retains each weight's direct influence on its postsynaptic neuron; indirect paths through other neurons are omitted. Recurrent connections remain active in the forward simulation (Section 3, Algorithm 1; Appendix A.6). Finally, all parameter groups are updated together. After computing the sensory, context, recurrent, and readout gradients from the same trial loss, we apply one Adam update per trial, using the parameter-group learning rates specified in Appendix A.3. All groups therefore adapt jointly throughout training (Section 3; Algorithm 1; Appendix A.3). **Q4. What is $h$, and how does it relate to $V_k$?** $V_k$ is the membrane potential of compartment $k$. The vector $h$ contains the somatic spike counts of the 40 neurons: $h_i$ is the number of spikes emitted by neuron $i$ during the 50–500 ms readout window. Thus, simulated somatic voltage trajectories generate the spikes that are counted in $h$ and passed to the action readout. Section 3 now defines both variables explicitly (Section 3, Eq. 3; Appendix A.3; Table S1). **Q5. What is $\bar g_S$?** This denotes the average sensory-weight gradient in the shared-learning experiment. To distinguish gradients from conductances, we now use $\mathcal G$ consistently in the main text and appendices: $$\overline{\mathcal G}_S =\tfrac12\left( \mathcal G_{S,\mathrm{apical}} +\mathcal G_{S,\mathrm{basal}} \right).$$ Each circuit computes its sensory gradient from its own task loss. The gradients are averaged before a single Adam update using a shared sensory optimizer state, and the updated sensory weights are applied to both circuits (Section 3; Appendix A.6).
AuthorsOpenReview ↗

Response to Weakness 2: Why does apical routing reduce dependence on the cue?

The core intuition is simple: when the explicit rule cue lands on the same basal dendrites as sensory input, optimization leans more heavily on the cue. When the cue instead arrives on active apical dendrites, calcium-mediated dendritic integration may modulate somatic firing in a way that keeps basal sensory updates useful even without the cue. Four lines of evidence support this account:

1. Active dendrites are essential—simplified models give the opposite result.

On the exact same task, both an abstract gain model and a three-compartment spiking model (under both BPTT and e-prop) perform worse without the cue when context is modulatory/apical (Γdep\Gamma_{\mathrm{dep}}=-0.2315, -0.1387, and -0.1365; new Appendix F, Fig. S15, Table S11). By contrast, an independent 20-compartment biophysical model with active channels reproduces the apical advantage (Γdep\Gamma_{\mathrm{dep}}=+0.0466). Making its dendrites passive cuts this gap to +0.0210, and removing only apical tuft calcium channels cuts it by approximately 80% to +0.0092, even though compartment separation, axial coupling, and NMDA synapses remain intact (Appendix F.3).

2. Training-period apical calcium conductance shapes later cue-free behavior.

In our detailed circuit, blocking apical slow-calcium conductance after normal training barely alters post-withdrawal error (ΔNMSE\Delta\mathrm{NMSE}=-0.00055), whereas blocking it during training raises post-withdrawal error by +0.014 even after we restore normal calcium conductance for testing (Fig. 4A–B, Table S9).

3. Apical routing keeps sensory learning gradients aligned with and without the cue.

Holding weights fixed at six checkpoints across training, we compared sensory-weight gradients (GS\mathcal G_S) computed with the cue on versus off (new Appendix G.2, Fig. S17). Apical routing shows higher cosine similarity at all six checkpoints (mean 0.709 vs. 0.646), meaning training with an apical cue pushes sensory weights along directions that still work once the cue is removed.

4. Feedforward sensory weights (SS) store this bias.

Swapping trained SS between apical and basal circuits transfers the cue-free error gap in both directions across all seed pairs (raising apical error from 0.145 to 0.197 and lowering basal error from 0.197 to 0.159), and swapping the weights associated with either the direction-encoding or the rule-encoding sensory channels contributes to this shift, whereas swapping recurrent weights RR has almost no effect (Fig. 3B–C, Table S3). Tying SS during training cuts the post-withdrawal error gap by 67% and the withdrawal-cost gap by 49% (Fig. 3E–F), and apical circuits preserve stronger target-action tuning upon cue removal (retention 0.704 vs. 0.654; new Appendix G.1, Fig. S16).

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

This paper investigates how the organization of dendritic inputs influence learned sensorimotor responses by trainng recurrent network models of multicompartment neurons. The authors show that apical and basal contextual inputs achieve similar performance but result in different dependencies on the redundant conextual input. The configuration of sensory weights and apical calcium conductances together contribute to the robutsness of the apecal network on context input. In a simulated robotic task, apical context network delivered stronger performance comapred to the basal network or context-free network. This performance gain depended on the persistent presence of context through execution.

Critical strengths

The paper raises a clear and relevant question: how does the dendritic location of contextual input affect how the network relies on contextual information? The authors designed detailed perturbation experiments to test the contribution of sensory weight configuration, training protocol, and calcium conductance on the superior performance of the apical network. Carefully deisgned control experiments make these comparisons rigorous.

Critical weaknesses
  1. Limited novelty: Compartment-specific dendritic gating (Yang et al., 2016), denritic modulation in representation learning (Wybo et al., 2023) and apical-basal integration of sensory and contextual information (Islah et al., 2025) have all been examined using modeling approaches. There is no major architectural or algorithmic innovation compared to previous works.

  2. Unclear computational mechanism: There is no clear explanation on why the dendritic location of the contextual input differentially affect learning. Nor is there insight on why sensory weight configuration and shared training protocol contribute to context dependence. It is hence unclear whether the findings reveal a general computational principle or are a result of specific task or model designs.

References: Yang, G. R., Murray, J. D., & Wang, X.-J. (2016). A dendritic disinhibitory circuit mechanism for pathway-specific gating. Nature Communications, 7, 12815 Wybo, W. A. M., et al. (2023). NMDA-driven dendritic modulation enables multitask representation learning in hierarchical sensory processing pathways. PNAS, 120(32), e2300558120. Islah, N., et al. (2025). Learning to combine top-down context and feed-forward representations under ambiguity with apical and basal dendrites. Cerebral Cortex, 35(6), bhaf134.

Questions for authors
  1. What distinguishes the findings in this paper from previous modeling studies on dendritic gating and integration?
  2. Can the authors provide an explanation on the why apical routing of contextual input yields an advantage on learning? Is it due to difference in the integration of context information, the propagation of gradient information or the learned representation of task variables?
  3. Can the robotic manipulation result generalize to more than one task? What is the key mechanism by which apical routing is beneficial in such tasks?
6 comments
AuthorsOpenReview ↗

Response to Question 1

Thanks for your thoughtful comments and for recognizing the relevance of our question and the rigor of our perturbation experiments. Our central finding is simple: where an extra rule cue arrives on the dendritic tree changes how well the sensory pathway learns to solve the task on its own. To address your questions, we tested simpler neuron models on our task (Appendix F), analyzed sensory gradients and learned tuning (Appendix G), and added a second robot task, Hand-insert (Appendix I), with all changes marked in blue.

Q1. Distinction from earlier models of dendritic gating and integration

Earlier dendritic models study basic context-dependent tasks in which the context cue stays on during execution. Motivated by recent in vivo neuroscience findings, we study a more advanced task involving cue withdrawal after learning, in which earlier simplified models fail and active dendrites become the core mechanism.

1. A more advanced task: post-learning cue withdrawal vs. basic online gating

Prior models (Yang et al., 2016; Wybo et al., 2023; Islah et al., 2025) focus on tasks in which the network always receives the context cue at test time to gate or combine inputs online. However, recent in vivo experiments uncovered a sharp split between learning and execution: inhibiting apical tufts impairs rule relearning while largely sparing established behavior (Maristany de las Casas et al., 2026). Capturing this split requires testing what the network retains after the cue is removed. In our task, sensory input already carries both direction and rule, while context repeats the rule. When the cue stays on, apical and basal networks perform nearly identically (0.053 vs. 0.051 NMSE), so basic cue-present evaluations cannot tell them apart. Only withdrawing the cue after training reveals that apical routing taught the sensory pathway to perform much better on its own (0.145 vs. 0.197 NMSE; Figs. 2A and 3).

2. Joint cross-pathway learning vs. fixed sensory weights

Because earlier models only needed context to modulate ongoing activity while the cue was present, they did not test how context location shapes sensory learning. In the spiking network of Wybo et al. (2023), for instance, feedforward sensory weights are held fixed, only context synapses learn, and dendritic membranes are passive aside from NMDA receptors. Our cue-withdrawal task requires jointly training sensory, context, recurrent, and readout weights in biophysical neuronal circuits so that apical activity during learning can shape basal sensory weights (Figs. 1C and 3–4).

3. Across neuronal abstraction levels: active biophysics captures the apical advantage missed by simplified models

When we run simplified models from prior work—an abstract gain model (Wybo et al., 2023) and a three-compartment spiking model (Capone et al., 2023)—on our cue-withdrawal task, apical/modulatory routing performs poorly without the cue and produces a pattern opposite to that of our main circuit (Appendix F; Fig. S15; Table S11). By contrast, an active 20-compartment biophysical model performs well and reproduces the apical advantage unless we make its dendrites passive or block tuft calcium channels, reducing the withdrawal-cost gap from 0.047 to 0.021 and 0.009, respectively. Simplified models are sufficient for basic cue-present gating, but active dendrites are required for this post-withdrawal learning regime.

AuthorsOpenReview ↗

Response to Question 2: Mechanism of the apical learning advantage

All three factors you asked about—context integration, gradient propagation, and learned representations—work together as one causal chain. In plain terms, routing the extra rule cue to basal dendrites makes the network lean heavily on that cue, whereas routing it to active apical dendrites helps the basal sensory pathway learn to solve the task on its own:

1. Active dendritic integration

Separating input locations alone does not create the apical advantage: both the abstract gain model and the three-compartment model (under both BPTT and e-prop) produce the opposite pattern, whereas an active 20-compartment biophysical model reproduces our main result (new Appendix F.1–F.2; Fig. S15; Table S11). Making dendrites passive or removing only apical tuft calcium channels in that 20-compartment model reduces the withdrawal-cost gap from 0.047 to 0.021 and 0.009, respectively, even though compartment separation, axial coupling, and NMDA synapses remain intact (new Appendix F.3; Fig. S15B,D). In our detailed apical-context circuits, blocking apical calcium after normal training barely changes cue-free error (a change of −0.00055 NMSE), whereas blocking it during training increases cue-free NMSE by 0.014 even after we restore calcium channels at test time (Fig. 4A–B). Apical calcium conductance during training therefore shapes the circuit’s learned ability to act without the cue, with effects that persist after conductance restoration.

2. Aligned sensory gradients

How does apical activity during training help basal sensory synapses learn? At six checkpoints across training, we held network weights fixed and compared the sensory-weight gradients computed with the cue present versus with the cue removed (new Appendix G.2; Fig. S17). Apical routing kept the cue-present and cue-free sensory gradients better aligned at every checkpoint (mean cosine similarity: 0.709 vs. 0.646 for basal routing). Put simply, while the cue is on during training, apical routing pushes sensory weights in directions that remain useful once the cue is removed.

3. Learned sensory weights

Because sensory input already encodes both direction and rule, better-aligned gradients train the feedforward sensory weights (S) to support behavior more independently. This directly explains why both of our sensory-weight interventions work (Figs. 3 and 4C). Swapping S after training transfers the learned behavior in both directions: putting basal-trained S into an apical network raises its cue-free error from 0.145 to 0.197 NMSE, while the reverse swap lowers basal error from 0.197 to 0.159 (Fig. 3B). Swapping S between intact and calcium-blocked runs similarly transfers the calcium-training effect (+0.053 and −0.023 NMSE; Fig. 4C). Conversely, sharing S during training averages both networks' sensory-weight gradients into a single update (Eq. 5), preventing their sensory weights from diverging and cutting the cue-free error gap by 67% (Fig. 3D–F).

4. Preserved target tuning

Because it learned a stronger standalone mapping through its sensory weights, the apical network preserves more of its target-action tuning when we remove the cue after training (retention: 0.704 vs. 0.654; directional similarity: 0.943 vs. 0.916; new Appendix G.1; Fig. S16). In sum, active apical calcium during training aligns sensory gradients, builds more self-sufficient sensory weights, and preserves target tuning without the cue.

AuthorsOpenReview ↗

Response to Question 3: Generalization across robotic tasks and the role of context

Yes—the apical advantage generalizes to a second manipulation task, Hand-insert. Our interpretation is that apical routing lets stage cues guide what the sensory pathway learns, while preserving its ability to support control after cue withdrawal.

1. Generalization to a second manipulation task

Following your suggestion, we tested Meta-World Hand-insert, where the arm must approach and grasp a block, carry and align it, and lower it into a slot—a very different action sequence from pushing a mug in CoffeePush (new Appendix I; Fig. S20; Table S12). Across three seeds per route, apical routing achieves higher success with the stage cue present and shows a smaller drop in success when we remove it:

TaskApical success with cue (%)Basal success with cue (%)Apical withdrawal cost (pp)Basal withdrawal cost (pp)
CoffeePush49.4045.841.373.46
Hand-insert41.6437.672.065.34

Importantly, apical routing still yields higher mean success even when we remove the cue at test time. In Hand-insert, apical controllers show smaller withdrawal costs than basal controllers. Their cue-free success rate (39.58%) exceeds those of both basal controllers evaluated without the cue (32.33%) and controllers trained without any cue (37.09%; Table S12). CoffeePush shows the same apical-routing benefit after cue removal: mean success is 48.03% for apical controllers, compared with 42.38% for basal controllers and 41.90% for controllers trained without the cue (Table S7).

2. Why apical routing helps in closed-loop control

In both tasks, sensory input gives the current 3D positions of the gripper, object, and goal, while the context cue uses trajectory history to label the current stage (such as reach, contact, or push). A linear classifier can already predict the stage from the current sensory frame alone with 97.7% accuracy in CoffeePush and 97.33% in Hand-insert (Appendices D.2 and I.2). Thus, just like the rule in our main task, stage information is mostly already in the sensory input, while the cue provides clean stage boundaries. Still, knowing the stage is not enough: within every stage, the arm must continuously read sensory positions to choose its exact movement direction and distance. As our weight-swapping and calcium interventions showed in the controlled task, routing the cue to apical dendrites lets the network use stage information while preserving strong sensory-driven control on basal dendrites.

3. Direct test during learning: apical cues help, basal cues interfere

Paired continuations in CoffeePush test this training effect directly (Fig. 5A,C; Appendix D.4; Table S6B). Starting from the same controller and optimizer state at epoch 20, we continued training two branches through epoch 50—one with the stage cue kept on and one with it removed. With the cue present at evaluation, retaining it during continued training changed success by 9.42 percentage points for apical routing and −2.13 percentage points for basal routing. The continuation benefit was most consistent when the cue remained available at evaluation. Overall, the exact same stage cue helps learning on apical dendrites while interfering with learning on basal dendrites.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Responses to Reviewers

Thanks for your thoughtful comments, Reviewers oPEY, GbWD, and 4x4C. In our main task, sensory input already tells the network both the stimulus direction and which rule to use, while an extra cue repeats the rule. Both networks perform similarly when this extra cue is present, but when we remove it after training, the network that received the cue on apical dendrites makes fewer errors than the one that received it on basal dendrites. In short, where the extra cue arrives changes how well the sensory pathway learns to do the task on its own, and active dendrites are key to this effect. We made four main revisions to address your questions and marked all changes in blue:

  1. Active dendrites are key (Reviewers oPEY, GbWD, 4x4C). Abstract gain and three-compartment models do not reproduce the apical advantage after cue withdrawal, while an active 20-compartment biophysical model does; tuft Ca blockade weakens this advantage (Appendix F).

  2. We examined what the sensory pathway learns (Reviewers oPEY, GbWD). Apical routing kept sensory gradients more aligned with and without the cue during training, and better preserved target-action tuning after cue removal (Appendix G).

  3. We tested when the result holds (Reviewer 4x4C). The apical advantage held across ~6–24 Hz firing rates and with longer training; when only the cue carried the rule, both networks failed as expected without it (Appendices B, H).

  4. We added a second robot task, Hand-insert (Reviewers oPEY, 4x4C). Removing the cue cut success by 2.06 percentage points for apical routing vs. 5.34 for basal, with smaller apical losses in all three seeds (Appendix I, Table S12).

Finally, in response to Reviewer GbWD, we rewrote Section 3 and Algorithm 1 (with full equations in Appendix A.6) to define every variable and show step by step how each trial's action error updates the sensory, context, recurrent, and readout weights together.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.