acceptodds
Withdrawn from ICLR 2027

Learning by Growing: Dendritic Development Shapes Task Representations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In biological brains, form and computation are intrinsically linked: dendritic growth changes both which inputs a neuron receives and how it integrates them. Standard artificial neural networks represent connections as numerical weights without modeling their physical structure. This raises a question for brain-inspired AI: can task experience guide physical growth to build useful representations? Here we train simulated biophysical neuron populations through task-guided, geometrically constrained dendritic growth and pruning. Synaptic strengths follow geometry rather than independent optimization, and a separately fitted linear readout predicts from population responses. Across synchrony and displacement tasks, growth builds predictive representations. Crucially, reassigning the same selected inputs among neurons reduces performance even when learned trees and contact strengths are preserved and the readout is retrained. This allocation advantage develops during learning. Learned populations also benefit more from combining neuronal responses than remapped populations do, supporting the value of complementary population features. More connections do not always help: in synchrony detection, active dendritic performance declines modestly at higher connectivity while a point-neuron baseline continues to improve. These findings connect biological development to representation learning, showing how task-guided growth establishes useful input organization beyond source selection. Ultimately, they reveal how intelligent representations are fundamentally rooted in their underlying physical morphology, providing a framework for investigating the formation of structural inductive biases.

Primary area:
applications to neuroscience & cognitive science
Keywords:
computational neuroscience, dendritic computation, structural plasticity, biophysical neuron models, population coding

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1easOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript studies whether dendritic growth can be used as a task-driven learning mechanism without backpropagation. The authors iteratively modify dendritic morphology, simulate the resulting neuronal responses, and evaluate each candidate structure using a refitted linear readout. Experiments are conducted on synchrony detection, direction classification, and displacement estimation. The results show that performance improves as morphology develops, that redistributing the selected inputs across neurons reduces performance, and that increasing connectivity does not always lead to better prediction. The manuscript further analyzes these observations through remapping experiments and comparisons among HH, Passive, and Point neuron implementations.

Critical strengths

The research question is interesting, and studying neuron-level learning through structural adaptation rather than conventional backpropagation is a reasonable and potentially useful direction. And the remapping experiments go beyond a simple before-and-after performance comparison and provide some analysis of how input organization affects the learned representation.

Critical weaknesses
  1. The main contributions are relatively basic and largely consistent with existing structural-learning knowledge. The manuscript shows that structural adaptation can improve performance, that input organization matters, and that denser connectivity is not always better, but it does not clearly establish what new insight is specifically enabled by dendritic computation.
  2. The current experiments do not sufficiently isolate which structural or biophysical factors are responsible for the observed gains. In addition, the Point model differs from the dendritic models in input representation and training procedure, which complicates the interpretation of the connectivity–performance comparison.
Questions for authors
  1. The main findings are broadly consistent with known structural-learning behavior. What new insight does the dendritic setting provide beyond showing that these effects also occur in a biophysical neuron model?
  2. The Point model differs from the dendritic models in both input representation and training procedure. How are these differences excluded as explanations for the different connectivity–performance trends?
  3. In the source-held remapping experiment, reassigning source identities to fixed contact sites changes the dendritic location and local integration context associated with each source. How can the effect of neuron-level allocation be separated from the effect of within-neuron dendritic placement?
  4. Morphological growth simultaneously changes connectivity, effective synaptic strength, and dendritic signal propagation. Although HH and Passive provide a partial comparison, which of these coupled factors is primarily responsible for the reported improvement?
Reviewer_1wUkOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This work models tree-structured neurons using the HH model and growable compartments, with connection weights determined by the distances between contact sites and signal sources. Its main contribution is to guide synaptic formation using task performance measured by a linear readout. The authors study three tasks and compare two types of remapping baselines. Their results show that performance declines when the connections between signal sources and dendrites are disrupted.

Critical strengths

1.The problem studied in this paper is interesting and well aligned with ICLR’s scope. 2.The three task settings are valuable. Although simple, they may provide a useful basis for comparison with corresponding neuroscience findings.

Critical weaknesses

1.The paper’s actual contribution is very limited and does not match the framing in its title and abstract. The title and abstract give the impression that the authors are the first to propose dendritic development as a learning method, yet the related work describes many similar studies. The main methodological contribution appears to be selecting each candidate dendritic growth step based on task performance. This is a greedy selection procedure and, in itself, does not seem novel. The presentation gives me the impression that prior work has been underemphasized, thereby overstating the paper’s novelty. 2.The paper is poorly written. Related work should not be deferred to the end of the paper. Many key methodological details are relegated to the appendix, making the paper difficult to understand without AI assistance. 3.The two remapping baselines provide little scientific insight. It is unsurprising that changing the inputs to a trained network would degrade its representational performance. Likewise, “more is better” does not seem to be a sufficiently compelling hypothesis to test. Despite the sophisticated mathematical modeling and appealing task settings, the two remapping baselines reduce the study to a toy model. The results offer neither a surprising finding nor theoretical insight.

Questions for authors

1.What are the substantive differences between this work and prior studies on dendritic growth? Do these differences constitute valuable innovations? 2.The two remapping baselines appear to provide little new insight. Rather than constructing a toy model, the study should be motivated by a concrete problem in neuroscience. 3.The methods section is very difficult to follow, especially for readers unfamiliar with dendritic growth research. Please explain how dendritic modeling is added to the classical HH model, and describe the simulation and evaluation procedures. The candidate actions, including what they are and how they are selected, should be explained in the main text.

Reviewer_xT9kOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Inspired by the complex branching structures of dendrites in the brain, this paper goes beyond training neural networks only through their weights by considering the morphology of the network. This not only allows the weights of connections to be learnt but gives freedom in how each neuron integrates its different inputs. They do this by evaluating the previously developed HelioX (lu et.al., 2026) on three different tasks, also performing several further evaluations and ablations. Their results show that reallocating the sources results in reduced performance across all three tasks, and that increasing the input connectivity budget doesn't always increase performance. Both of these results are suggested to support the motivating idea that dendrites allow learning not only through the strength of their connections, but also the way in which they are connected. Implying that moving beyond point neurons, by introducing additional degrees by considering network geometry, may improve learning

Critical strengths
  • The question is very interesting and well motivated. The specific approach and questions asked are of interest to both the computation neuroscience community and beyond.

  • The overall approach and conclusions drawn from the results are reasonable and sound.

  • The HelioX framework is well designed and suited to the questions posed.

  • The paper is readable and easy to understand

Critical weaknesses
  • Limited contribution. There are no methodological contributions in the paper, with the HelioX framework being an already published result. Thus the contributions are solely in the empirical evaluation, which as discussed below we do not believe are sufficient.

  • The results are unsurprising. They cite three main contributions in the paper: That performance increased across the course of learning, that input reallocation reduced performance and finally that increasing the capacity budget did not always lead to improved performance. Given that the framework proposes new connections, evaluates the performance of each then chooses the best performing option including no change, it is hard to see why performance would not improve and thus why this is an interesting result. Given the hypothesis of the paper is that by including geometry in their framework, one can improve learning, through the additional freedom in choosing how to combine inputs, it is of course encouraging that reallocation reduces performance supporting this hypothesis. However, this result in isolation isn't too interesting especially given that the point neuron control performed just as well if not better on most tasks and that (as will be discussed later) the experiments were too simple to exploit any of the possible interesting advantages of morphology.

  • The final contribution on the affect of the weight capacity seems incomplete. The result and question in itself is intriguing however this was only observed in one of the three tasks. With this task being the only one which included 'source distractors', it would thus be interesting why only this task saw reduced performance and why.

  • Disconnect between the motivation and experiment design. The introduction argues that physical geometry allows neurons to combine inputs and perform dendritic preprocessing before signals reach the soma, capabilities beyond those of a point neuron. However, the chosen experiments do not actually require or exploit these advantages. For example, in the direction task, the learned morphologies rarely combine inputs before reaching the soma (see Figure 2e or Section 4.3) and the solution obtained resembles simple unimodal tuning curves (see fig 3g) like what we observe in the V1 cortex, where understanding the solution in this way makes the reallocation ablation result unsuprising and missing the point. Consequently, the morphological learning here seems to just act as an overly complex method for standard weight updates, which explains why the point neuron control performs just as well.

  • The paper is missing any intuition or analysis on why we see the observed results, e.g. for example how the learnt structures utilised geometry in their solution and thus why the performance as hurt by the reallocation (thus a counter argument to the previous comment), beyond through factors which would be similarly seen in the point neuron approach.

  • Furthermore, some intuition either through an example from the experiments or a toy model of how the learnt representation can benefit from its morphology beyond factors that can be just explained by weights would be appreciated. In particular a motivating example of where morphological preprocessing solves a problem that purely weight-based learning cannot. This should both be possible and very illuminating and is exactly why we think the research question is well justified.

Questions for authors
  • We request the authors to highlight examples where the learnt solutions do in fact make use of their morphology and provide some heuristic as to what advantage is gained by combining these inputs in such a way

  • We request further analysis on the reduction in performance seen in task 1 with increasing weight budget, ideally demonstrating it is not just an artifact of the source distractors or due to weaker orientation selectivity in the neurons, i.e. overly broad tuning curve widths

  • It would be of interest to have a further source reallocation test, where reshuffling only occurs within the sources used by each neuron. We suggest this as we understand the central premise of why complex dendritic structure is to allow for pre-processing of the input signals before they even reach the soma, thus one would hope keeping the inputs to each neuron the same but reshuffling the way in which they are processed should reduce performance, as ideally the algorithm should not only identify the best sources to respond to but the way to integrate them. We argue this is a stronger test more reflective of the intended goal than either the source held or full graph tests currently used.

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

Related papers

Discussion (0)

Sign in to comment.