acceptodds
Under review as a conference paper at ICLR 2027

SynMorph: Learning to Build Task-Adaptive Spatiotemporal Models from Synaptic Primitives

Abstract

In real-world spatiotemporal learning applications, a single observation system often involves diverse learning tasks with heterogeneous objectives and data patterns. Researchers still need to manually design task-specific models according to data characteristics, even when tasks share the same spatiotemporal observations but require different input-output mappings, resulting in considerable engineering efforts. Existing architectures addressing task diversity mainly rely on fixed-structure adaptation, such as transfer learning, or architecture search strategies that lack structural inheritance and composition. Inspired by synaptic diversity in biological brains, we propose Synaptic Morphogenesis (SynMorph), a task-adaptive framework that learns to construct spatiotemporal models from heterogeneous synaptic primitives according to task-specific data patterns. Specifically, SynMorph first generates task encoding by jointly characterizing task types and spatiotemporal patterns. It then constructs a neuron-level meta-synaptic pool inspired by diverse synaptic mechanisms and adaptively selects synaptic primitives to compose task-specific networks through a shared Deep Q-Network (DQN) controller. The decision process is guided by network states and an overall reward function balancing model complexity and prediction performance. Finally, the composed architecture is trained end-to-end for the target task. Extensive experiments on diverse datasets involving regression and classification tasks demonstrate the effectiveness and interpretability of SynMorph, achieving up to 7.0% improvement over the strongest existing approaches on environment datasets. SynMorph provides a neuro-inspired perspective for task-adaptive spatiotemporal learning by uncovering the potential dependencies among data patterns, task requirements, and model architectures, which can be viewed as `the model of model'.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatiotemporal learning, Task-adaptive adaptation, Neuro-inspired deep learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_RGG3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Synaptic Morphogenesis, a biological inspired, task-adaptive framework that can select its components through a shared DQN to face different kind of tasks.

Critical strengths

The method incorporates the concept of biological neuronal modulation, enabling adaptive adjustment of network components and adaptation to entirely different types of data, such as traffic and EEG data.

Critical weaknesses

1: Task types need to be provided for this framework, yet the model not achieve the SOTA. Facing different kind of tasks, the task types must be clear and provided to the framework. Meanwhile, according to Table 1, SynMorph only achieves the best performances in classification tasks and failed in regression tasks. In Table 2, SynMorph failed in CHI R. Consequently, the advantages of the proposed method are not outstanding, it’s easier to choose different SOTA models for each task, because the task type is provided.

2: Clearer emphasis on novelty. Using reinforcement learning to select neural network architectures is not a new idea, as it was already explored in MetaQNN [1]. The authors should discuss the differences and better emphasize their novelty, such as what advantages synaptic mechanism provided comparing to the traditional RL method.

[1] Bowen Baker, et.al. Designing Neural Network Architectures using Reinforcement Learning, ICLR2017

Questions for authors

1: The training and inference process for the SynMorph I am not entirely clear on this point: does the framework use a task-type encoder during training, but a fixed network architecture for each specific task during inference? If my understanding is incorrect—for example, if the framework can automatically adapt its architecture to the characteristics of different datasets without relying on prior knowledge of the task type—I would consider raising my score.

2: The meaning of Figure 2 and Figure 3 It’s hard to understand these figures, what the authors want to emphasize in these Figures?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RGG3 part I

W1. Concern on tasks that SynMorph doesn’t achieve SOTA

Thanks for your valuable comments. Our experiments primarily aim to validate the effectiveness of SynMorph under heterogeneous task settings. Therefore, its strong classification performance also demonstrates the effectiveness of our approach. Regarding the tasks where SynMorph does not achieve SOTA performance, we believe an important reason is that existing specialized models often incorporate extensive domain-specific prior knowledge into their architectural designs, e.g., domain-informed spatial fusion in DeepAir, bidirectional diffusion-based recurrent modeling in DCRNN, and EEG-specific spectral-power extraction in ShallowNet. In contrast, SynMorph aims to construct neural architectures from scratch, primarily based on observable data and task information, without relying on additional domain-specific modules. Therefore, these approaches differ in their research objectives and reliance on domain priors. Our goal is to learn architecture-construction strategies autonomously using only observable inputs and necessary task information. Our results show that SynMorph achieves SOTA performance on 50% of the evaluated tasks and remains competitive on the others. These results suggest that, even without carefully designed domain-specific mechanisms, SynMorph can maintain strong performance across heterogeneous tasks, demonstrating its data-driven architecture construction capability and task adaptiveness.

W2. Concern of novelty and comparison with MetaQNN.

Thanks for your reflective comments. Unlike MetaQNN, which uses Q-learning to sequentially select conventional CNN layers and their configurations, SynMorph progressively constructs neural architectures from scratch according to task requirements and data characteristics. Moreover, we construct a meta-synaptic pool where each unit constitutes a sub-neural architecture designed to achieve specific functions like spatial integration and information projection inspired by the morphological and functional diversity of neural synapses, and our RL controller performs feedback-driven regulation of this bottom-up structural growth. At each decision step, the current network state and data-task descriptors jointly guide synaptic primitive selection. As new mini-batches arrive, the data-dependent descriptors can be updated, allowing structural decisions to respond to changing input characteristics. Meanwhile, feedback from prediction performance, together with structural considerations, provides rewards to progressively optimize the architecture-construction policy.

Our existing results provide complementary evidence for this design. In Table 4, replacing learned primitive selection with random composition increases KnowAir MAE from 34.28 to 50.98 and decreases SIP classification accuracy from 96.89% to 85.24%, demonstrating the importance of learned functional composition. Removing individual primitive categories further degrades performance, confirming the value of functionally heterogeneous structural choices. Moreover, reinitializing the controller for each task reduces SIP classification accuracy from 96.89% to 91.52%, supporting the benefit of retaining construction knowledge across tasks. Figures 2 and 7 further illustrate task-dependent projection preferences and functional composition patterns, providing structural evidence of customized model construction.

Following your suggestions, we will expand the discussion of MetaQNN and clarify that SynMorph's novelty lies in task-conditioned composition of heterogeneous functional primitives and reusable cross-task architecture-construction policies, rather than the reinforcement learning algorithm itself.

AuthorsOpenReview ↗

Response to Reviewer RGG3 part II

Q1. Confusion of flexibility of SynMorph

Thanks for your comments. We do not assign a fixed architecture to classification or regression tasks. Instead, SynMorph automatically identifies and constructs a suitable neural architecture according to the input data, allowing the resulting architecture to fit the underlying data patterns. We would like to clarify the following two points:

  1. Task-type encoding is necessary, because the task type must be specified to determine the type and content of the model output, which has been confirmed in Table 4.

  2. Our core contribution lies in the data-sensitive nature of architecture construction. Even without domain-specific prior knowledge about the data, SynMorph can adaptively construct and adjust its architecture through data-pattern encoding and task encoding. Specifically, the state represents the current network structure together with data and task characteristics, while the reward provides feedback on the learning process, reflecting how well the current architecture fits the observed data. Through this feedback-driven process, SynMorph progressively regulates its architecture and learns a suitable structural configuration until the construction process converges.

Q2. The meaning of Figure 2 and Figure 3

Figure 2 illustrates that classification and regression tasks have different neural architecture preferences, which has been verified and corresponding to the design of SynMorph. Even when regression and classification tasks share the same input dataset, SynMorph exhibits distinct preferences for synaptic primitives, leading to different network architectures. This visualization highlights the interpretability of SynMorph's task-adaptive structural decisions.

Figure 3 illustrates the sequential synaptic selection process through a concrete example and helps better understand strategy construction process of our methodology. This example demonstrates how reward feedback guides the progressive construction of task-adaptive architectures. We appreciate you again for highlighting this ambiguity.

I thank the authors for their further clarifications and responses. Overall, these responses have partially addressed my concerns. However, several issues remain unclear to me, and I believe they cannot be adequately resolved through textual explanations alone.

  1. Regarding the task encoder: The authors emphasize that one of SynMorph's key advantages is the data-sensitive nature of its architecture construction process. However, removing the task encoder leads to a substantial performance degradation, which seems somewhat inconsistent with the claimed data sensitivity. Furthermore, even without the task encoder, task-related information is still provided through the loss function. This may explain why performance does not completely collapse after removing the task encoder, rather than demonstrating an inherent advantage of the architecture construction mechanism itself.
  2. Regarding novelty: I am not convinced that the construction of heterogeneous neural architectures constitutes a novel contribution, as this idea has already been explored in prior research.
  3. Regarding Figures 2 and 3: I appreciate the authors' further clarification, which has resolved my confusion on this point. However, I believe these figures and their accompanying explanations should be integrated with Figure 1 to provide a more coherent presentation of the proposed methodology, rather than being placed at the end of the paper. Therefore, I have decided to maintain my original score.
AuthorsOpenReview ↗

Response to the follow-up comments of Reviewer RGG3

Thanks for your further discussions on our work. We hope the following 3 points can one-to-one address your updated concerns:

  1. Removing the Task Encoder means eliminating both the Task-Type Encoding and the Spatial/Temporal Descriptors, which jointly characterize task requirements and data patterns. Therefore, removing the entire Task Encoder eliminates not only task-type information but also data-dependent structural features, leading to performance degradation. To further investigate the data sensitivity of SynMorph, we conducted an additional ablation experiment on SIP by removing only the Task-Type Encoding while retaining the Spatial/Temporal Descriptors, i.e., SynMorph-Task type (MAE and Acc reported): SIP R:13.11 SIP C: 95.82. This experiment demonstrates that only data-pattern information can still effectively guide adaptive architecture construction without explicit task-type encoding.

  2. Our core novelty lies in resolving the problem of constructing learnable neural architecture from scratch given task type and datasets. To resolve this, we establish a series of neural functional structures inspired by morphological diversity, which allows the complexity and flexibility of data flowing, and then propose a reinforcement based neural structure connection strategy to form the neural architecture from scratch. We believe that such a mechanism, integrating fine-grained functional composition with task-data-driven, feedback-guided architecture construction, has not been exactly explored in prior work.

  3. Figure 2 illustrates the task-adaptive structural preferences learned by SynMorph across different tasks and serves as a further interpretation of our experimental results. Therefore, we believe it is more appropriate to retain it in the experimental analysis section rather than integrate it into the framework figure. Figure 3 illustrates the architecture construction and decision-making process and could indeed be integrated with Figure 1. However, since Figure 1 already contains substantial information, we present Figure 3 separately to avoid overcrowding and improve readability. Based on your suggestion, we can remove Figure 2 into the Appendix of our manuscript or merge Figure 2 with Figure 3 to make it more compact.

With the author–reviewer discussion period ending in about one day, we would appreciate your follow-up on whether our further responses dispel your concerns. We are happy to incorporate all these discussions and results into our revised manuscript to improve the overall quality of the manuscript.

The authors

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

SynMorph builds separate regression and classification networks for shared spatiotemporal inputs. A shared DQN controller adds modules one by one based on task type, spatial/periodic cues, and current structure, scores candidates using loss, depth, and squared gradient norm, and updates its policy from score changes before final training. Tested on five datasets (air quality, traffic, EEG), it improves Knowair accuracy and includes ablations on modules, task encoding, reuse, and rewards.

Critical strengths
  1. A useful setting with competitive results. Evaluation covers five datasets across three application areas (Tables 1–3). On Knowair, classification accuracy improves from 38.99% for AdaShare, the strongest classification baseline, to 41.70%, a gain of 2.71 percentage points (Table 1).

  2. Useful ablations and initial robustness checks. Table 4 ablates primitive groups, task encoding, controller reuse, reward terms, and random/greedy selection, alongside 5-run variance (Tables 1, 7–8) and SIP sensitivity checks (Table 6).

Critical weaknesses
  1. Section 1 and Appendix A.1 overstate how different this is from transferable NAS like CATCH (context-conditioned RL) or MetaD2A (dataset-conditioned generation). Neither uses this spatiotemporal setup, and Table 4 doesn't show what the shared policy actually adds over comparable transferable search. A direct comparison would help.
  2. The efficiency claim for the reward isn't well supported. Scaling the loss by c scales Eq. (12)'s squared-gradient term by c² unless parameters or reward weights change (and relative rewards still depend on it). Table 6 shows similar SIP scores at 0.8/1.0/1.2 without ranking analysis, Table 4's depth/gradient terms are only proxies, and Appendix A.2 reports +21% time and −30% space with no baseline, metric, or absolute numbers.
  3. The ablations don't isolate what the descriptors add beyond task type (Section 3.2, Eqs. 3–7). Table 4 drops all task encoding (including task type), spatial/temporal ablations drop primitive families instead of descriptors, and it's unclear if paired tasks even get different descriptors. Resets (Table 4), task order (Table 6), and Figure 7 support reuse, not descriptor gains.
Questions for authors
  1. Under Eqs. (4)–(6), do paired tasks on the same inputs get identical descriptors, or can task-specific parameters/updates separate them? Could you run paired searches (same controller, pool, budgets) that keep task-type encoding but replace spatial/temporal descriptors with matching constants?

  2. How does SynMorph’s shared policy differ from CATCH and MetaD2A, and which existing result isolates the benefit of that difference?

  3. Do you normalize or unscale Eq. (12)’s gradients? Do Table 6’s loss-scaling tests rerun full searches? Please specify the reference, space metric, and absolute values for the 21% time increase and 30% space reduction. Does time include controller training, search, and final training?

2 comments
AuthorsOpenReview ↗

Response to Reviewer VKNB part I

W1Q2. Comparison with NAS.

Thanks for your concerns. SynMorph differs in its fine-grained, data-customized construction mechanism. Specifically, CATCH transfers search knowledge through context-conditioned RL, while MetaD2A generates candidate architectures from dataset representations. SynMorph progressively constructs neural architectures from scratch according to task requirements and data characteristics. Moreover, we construct a meta-synaptic pool where each unit constitutes a sub-neural architecture designed to achieve specific functions like spatial integration and information projection inspired by the morphological and functional diversity of neural synapses, and our RL controller performs feedback-driven regulation of this bottom-up structural growth. At each decision step, the current network state and data-task descriptors jointly guide synaptic primitive selection. As new mini-batches arrive, the data-dependent descriptors can be updated, allowing structural decisions to respond to changing input characteristics. Meanwhile, feedback from prediction performance, together with structural considerations, provides rewards to progressively optimize the architecture-construction policy.

Our existing results provide complementary evidence for this design. In Table 4, replacing learned primitive selection with random composition increases KnowAir MAE from 34.28 to 50.98 and decreases SIP classification accuracy from 96.89% to 85.24%, demonstrating the importance of learned functional composition. Removing individual primitive categories further degrades performance, confirming the value of functionally heterogeneous structural choices. Moreover, reinitializing the controller for each task reduces SIP classification accuracy from 96.89% to 91.52%, supporting the benefit of retaining construction knowledge across tasks. Figures 2 and 7 further illustrate task-dependent projection preferences and functional composition patterns, providing structural evidence of customized model construction. We will clarify this distinction and better connect the existing results to our methodological contributions.

W2Q3. The reward’s efficiency interpretation and concern of time increase and space reduction.

Thank you for this rigorous observation. The loss-scaling tests do rerun full searches. Following your suggestion, our additional analysis of ranking shows that the Top-1 selected primitives remain the same across the evaluated loss scales (0.8, 1.0, and 1.2). This provides direct empirical evidence that moderate loss scaling does not alter the preferred structural decisions in the evaluated cases. Consistently, Table 6 shows stable final performance, with SIP MAE ranging from 12.878 to 12.994 and ACC from 96.80% to 96.87%. Furthermore, the gradient term is intended to characterize optimization dynamics, whose effectiveness is supported by Table 4. In order to remove the effect of loss scaling, we changed the reward function to $ R_ n=- \frac {\eta^{d({\theta}_ t^{(n+1)})}\lambda {\nabla{\theta}_ t^{(n+1)}}}{\mathcal{L}_ t^{(n+1)}/\sigma_ {out}}+\frac{\eta^{d({\theta}_ t^{(n)})}\lambda{\nabla{\theta}_ t^{(n)}}}{\mathcal{L}_ t^{(n)}/\sigma_ {out}}$, where the computation of gradient $ \nabla{\theta}_ t $ in Eq.(12) also removes the square term, i.e., $ \nabla{\theta}_ t=\frac{1}{d({\theta}_ t)}\sum_ {n=1}^{d({\theta}_ t)}||\nabla s_ {t_ n}||_ 2$. After making such modifications to the reward function, SynMorph still works across different datasets. Specifically, the results are shown below:

Knowair R:34.33 Knowair C:41.75

CHI R:1.12 CHI C: 97.31

SIP R:12.86 SIP C:96.89

HGD R: 2.96 HGD C:83.10

2A R:0.74 2A C:75.17

More updated results like ablations will be reported in the accepted version due to time limit.

Moreover, we specify that the reported ~21% time increase and ~30% space reduction are measured against directly training the final architecture selected by SynMorph on SIP with the same total number of training epochs. To be specific, the reference time is about 508 minutes, and SynMorph, including the whole process, uses about 615 minutes due to its dynamic construction and optimization process. For space consumption, we adopt an average cumulative GPU memory footprint, defined as the average of GPU memory usage across all training epochs, i.e., $ A=\frac{1}{E}\sum_ {e=1}^{E} M_ {e}$, where $ M_{e} $ denotes the GPU memory usage at epoch e. Since SynMorph progressively constructs its architecture rather than training the complete architecture throughout, its cumulative memory footprint is approximately 30% lower than that of the fixed architecture under the same epoch budget. Specifically, the reference space cost is about 3440 MB, and SynMorph costs about 2397 MB.

AuthorsOpenReview ↗

Response to Reviewer VKNB part II

W3Q1. Concern of spatial and temporal descriptors.

Thank you for the reflective comments. First, we clarify that the spatial and temporal descriptors are not fixed representations precomputed from the entire dataset. Instead, they are dynamically generated from the current input mini-batch during architecture construction.

Although paired regression and classification tasks share the same input dataset, they may sample different mini-batches during their respective architecture construction processes. Consequently, their spatial and temporal descriptors may vary across tasks and structural decision steps. In our task encoding design, task-type encoding distinguishes different learning objectives, while spatial and temporal descriptors provide complementary, data-dependent structural information. Together, they modulate synaptic primitive selection through Eq. (7).

To address ablation limitation, we conduct additional ablation experiments for each descriptor. Specifically, while retaining the original task-type encoding, we separately replace the spatial or temporal descriptor with a fixed, data-independent vector of the same dimensionality, keeping all other settings unchanged. These experiments allow us to evaluate the individual contributions of the descriptors beyond task-type information. The results are presented below, which demonstrates the importance of the descriptors:

SynMorph-spatial descriptor:

Knowair R:34.85 Knowair C:39.93

SIP R:12.97 SIP C:96.25

HGD R: 3.01 HGD C:81.91

SynMorph-temporal descriptor:

Knowair R:34.66 Knowair C:39.50

SIP R:13.06 SIP C:96.41

HGD R: 3.09 HGD C:82.20

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

This paper presents SynMorph, a task-conditioned architecture construction framework for spatiotemporal learning. The proposed approach aims to build customized neural architectures for diverse tasks (e.g., regression and classification) sharing the same spatiotemporal input space. SynMorph comprises three main components:

  • Task Encoding: Combines a one-hot task identifier, a spatial topology embedding derived via a localized graph propagation step, and a temporal periodicity descriptor derived via FFT peak frequency and magnitude.
  • Meta-Synaptic Pool:A discrete set of operational primitives grouped into four neuroscience-inspired categories: spatial integration, temporal dynamics, information projection, and neural modulation.
  • Task-Conditioned DQN Controller: A shared Deep Q-Network that sequentially selects operations from the pool to construct a sequential neural architecture, guided by a multi-objective reward function incorporating task loss, model depth, and average gradient norm.

The method is evaluated across three application domains, air quality (Knowair), traffic forecasting (CHI, SIP), and EEG decoding (HGD, BCI 2A).

Critical strengths
  • Addressing multi-task spatiotemporal modeling by learning a reusable architecture-generation policy (rather than independent search per task) is a practically relevant and interesting direction.

  • The empirical evaluation spans diverse data modalities, covering both regression and classification settings.

  • The authors provide detailed ablation and sensitivity studies, demonstrating the impact of each module (task encoding, primitive categories, DQN vs. greedy selection, and reward terms).

Critical weaknesses
  • The paper heavily frames the methodology as "synaptic morphogenesis" and biological diversity. However, computationally, the framework is a conventional RL-based neural architecture rearch. The "synaptic primitives" are standard deep learning building blocks: 1D/2D convolutions, GCN layers, etc. Wrapping standard deep learning operations in neurobiological terminology (e.g., "divergent projection synapse", "receptive-field synapse", "stochastic modulation synapse") does not change their functional reality and obfuscates the technical contributions.
  • Conditioning an RL search controller (e.g., RNN or DQN) on task/data meta-features is well-established in meta-NAS and multi-task architecture search. Combining simple FFT frequency peaks, degree/topology features, and a one-hot vector via Hadamard product (Eq. 7) seems to be an incremental design choice.
  • On the primary real-world benchmark tasks (regression for air quality and traffic), SynMorph is frequently outperformed by existing task-specific models. SynMorph claims superiority based on AvgRank across regression and classification. However, the classification tasks appear artificially constructed by discretizing continuous variables into quantile bins. Demonstrating higher rank on an artificially binned classification task does not compensate for significant regression degradation compared to state-of-the-art baselines.
Questions for authors

Please refer to the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer oaee

W1. The contribution clarification of methodology “synaptic morphogenesis”, “biological diversity”, as well as formulated “synaptic primitives”.

Thanks for your constructive comments. Our main contribution lies in constructing various structures by simulating the morphological diversity of neural synapses, enabling them to have different local structural operations. This allows for adaptive local neural structure selection and learning architecture construction for data and tasks. Specifically, Spatial Integration allows task-customized convolution aggregation with data-task encoding; Temporal Dynamics enables task-adaptive modeling of temporal dependencies through dynamic integration and delayed information transmission. Information Projection constructs an adaptive neural structure to propagate information in a selective manner.. Neural Modulation enables task-customized pathway regulation through stochastic activation control and adaptive information retention.

Our novelty does not rely on introducing entirely new low-level operators, but on making these distinct functional mechanisms available for task-conditioned structural composition.

W2. Distinguish RL search controller against meta-NAS and multi-task architecture search. The incremental contribution of combining FFT frequency peaks, degree/topology features, and a one-hot vector in Eq.(7).

Thanks for your reflective comments. Rather than searching for a subnetwork within a predefined architecture like NAS, SynMorph progressively constructs neural architectures from scratch according to task requirements and data characteristics. Moreover, we construct a meta-synaptic pool where each unit constitutes a sub-neural architecture designed to achieve specific functions like spatial integration and information projection inspired by the morphological and functional diversity of neural synapses, and our RL controller performs feedback-driven regulation of this bottom-up structural growth. At each decision step, the current network state and data-task descriptors jointly guide synaptic primitive selection. As new mini-batches arrive, the data-dependent descriptors can be updated, allowing structural decisions to respond to changing input characteristics. Meanwhile, feedback from prediction performance, together with structural considerations, provides rewards to progressively optimize the architecture-construction policy.

Our existing results provide complementary evidence for this design. In Table 4, replacing learned primitive selection with random composition increases KnowAir MAE from 34.28 to 50.98 and decreases SIP classification accuracy from 96.89% to 85.24%, demonstrating the importance of learned functional composition. Removing individual primitive categories further degrades performance, confirming the value of functionally heterogeneous structural choices. Moreover, reinitializing the controller for each task reduces SIP classification accuracy from 96.89% to 91.52%, supporting the benefit of retaining construction knowledge across tasks. Figures 2 and 7 further illustrate task-dependent projection preferences and functional composition patterns, providing structural evidence of customized model construction.

Therefore, Eq. (7) describes the representation method of the state in the learning architecture, but it is not the core contribution of our entire work.

W3. Inferior performances on regression tasks

We want to clarify that, 1. Classification tasks are motivated by real-world applications, e.g., air quality management requires air quality classification based on multi-source observations, while urban traffic management relies on traffic condition classification to support practical decision-making.

  1. Our experiments primarily aim to validate the effectiveness of SynMorph under heterogeneous task settings. Therefore, its strong classification performance also demonstrates the effectiveness of our approach. Regarding the tasks where SynMorph does not achieve SOTA performance, we believe an important reason is that existing specialized models often incorporate extensive domain-specific prior knowledge into their architectural designs. In contrast, SynMorph aims to construct neural architectures from scratch, primarily based on observable data and task information, without relying on additional domain-specific modules. Therefore, these approaches differ in their research objectives and reliance on domain priors. Our goal is to learn architecture-construction strategies autonomously using only observable inputs and necessary task information.

Our results show that SynMorph achieves SOTA performance on 50% of the evaluated tasks and remains competitive on the others. These results suggest that, even without carefully designed domain-specific mechanisms, SynMorph can maintain strong performance across heterogeneous tasks, demonstrating its data-driven architecture construction capability and task adaptiveness.

Thank you for your response. I really appreciate the authors' efforts and do think this paper's motivation is interesting. However,

  • Regarding W1: The authors stated, "Our main contribution lies in constructing various structures by simulating the morphological diversity of neural synapses, enabling them to have different local structural operations." However, as noted in my initial review, the paper wraps NAS concepts in heavy biological terminology. This is primarily a severe clarity and presentation issue, not a novelty issue, yet the rebuttal fails to directly address this concern.

  • Regarding W3: The authors did not provide any further discussion or additional experimental results regarding the regression tasks.

Therefore, I will maintain my initial rating.

AuthorsOpenReview ↗

Response to the follow-up comments of Reviewer oaee

Regarding W1

Thanks for further raising this overstated issue. We have toned down the description of a one-to-one correspondence between artificial neural network modules and biological neuronal structures. Instead, we emphasize that we construct a computational unit pool consisting of multiple fundamental deep learning modules, which are categorized into four groups according to their functions. Inspired by the morphological diversity of biological neurons, these computational units can be adaptively selected and composed based on task-data patterns, current network states, and model performance feedback, enabling the progressive construction of a complete neural architecture from scratch. Now, we have accepted your suggestion and correspondingly revised the detailed method description in Sec.3.3 of Methodology and the new manuscript can be seen in the new pdf file:

Old Methodology Part: Inspired by morphological diversity of biological synapses, we construct a function-informed meta-synaptic pool, where heterogeneous synaptic primitives provide complementary computational functions for spatiotemporal information processing. The proposed pool consists of following four types of synaptic primitives:

New text: Inspired by the morphological diversity of biological neurons, we construct a meta-synaptic computational unit pool consisting of multiple fundamental deep learning modules which are categorized into four groups according to their functions:

Regarding W3

Regarding the regression tasks where SynMorph does not achieve SOTA performance, we believe an important reason is that existing specialized models often incorporate extensive domain-specific prior knowledge into their architectural designs, e.g., domain-informed spatial fusion in DeepAir, bidirectional diffusion-based recurrent modeling in DCRNN, wind-informed pollutant transport modeling with geography-constrained graph construction in PM2.5-GNN. We conduct the experiments on baselines mentioned above removing extensive domain-specific prior knowledge and the results are shown below:

Knowair (MAE reported): DeepAir Regression: 37.71; PM2.5GNN Regression:38.95

SIP (MAE reported): DCRNN Regression: 13.676.

From the results above, we can see that the performances of strong baselines drop worse than SynMorph without extensive domain-specific prior knowledge, which demonstrates the effectiveness of our adaptive SynMorph.

With the author–reviewer discussion period ending in about one day, we would appreciate your follow-up on whether our further responses dispel your concerns. We are happy to incorporate all these discussions and results into our revised manuscript to improve the overall quality of the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.