acceptodds
Under review as a conference paper at ICLR 2027

Building Network Functions from Signed Feedforward Loops

Abstract

Connectivity shapes computation by determining how signals propagate and interact within a network. Compared with analyses of individual connections, the feedforward loop (FFL) provides a minimal multipath structure in which direct and indirect pathways converge on the same target through a relay node. FFLs have been widely studied for their computational capacity, and their excitatory–inhibitory (E-I) configurations are known to determine distinct dynamical regimes. However, the role of E-I organization in the functional specialization of FFLs remains unclear. Here, we systematically analyze the dynamics of four E-I configurations of the FFL and characterize how direct and indirect pathways interact to shape computation. Under defined input and background conditions, these dynamics support distinct computational functions, including 1) response initiation, 2) response switching, 3) stopping of ongoing responses, and 4) context-dependent release. We then extend motif regularization with a differentiable signed-motif loss that increases the proportion of selected FFLs in larger networks, translating local computational mechanisms into structural priors for network learning. Behavioral evaluations in MiniHack show that these motif-enriched networks exhibit task-dependent computational biases consistent with their local functional specialization, and that a learned scheduler can coordinate components shaped by different motif priors to improve performance in more complex environments. Together, these results show how E-I organization shapes the functional specialization of FFLs and how local signed pathway dynamics can be transformed into reusable computational building blocks for constructing and composing functions.

Primary area:
learning on time series and dynamical systems
Keywords:
Signed feedforward loops, Excitatory–inhibitory organization, Neural dynamics, Structural priors, Network computation

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

This paper studies how different signed three-node feedforward loops (FFLs) support different computational functions. The authors analyze four signed FFLs with different configuration of E-I connections. They conduct experiments on MiniHack tasks show that different motif priors can improve learning performance by optimizing a joint loss function combined with PPO and motif regularizer.

Critical strengths
  1. The paper studies an interesting question about how local connectivity structures influence neural network computation. Connecting simple circuit dynamics with the design of larger RNNs is a meaningful direction.

  2. The paper is generally well organized. The authors evaluate the proposed structural priors on several reinforcement-learning tasks.

Critical weaknesses
  1. The proposed three-node model has a strictly feedforward structure, and the stability result (Proposition 1) follows directly from its triangular Jacobian.
  2. The authors linearize the dynamics around an equilibrium, but do not explicitly characterize the equilibrium in the main text. Actually this equilibrium can be explicitly characterized with the paper's assumption of constant input. In addition, the theoretical result of proposition 1 is derived, but if and how it will help to understand the four input-dependent computations are not analyzed. I suggest doing further analysis on how the linearized system behave in the four cases.
  3. The authors show that individual FFLs exhibit certain response patterns and that motif-regularized RNNs perform better on corresponding tasks. But they do not show that the trained RNNs actually use these local mechanisms. It is unclear whether the improvements result from the proposed motif functions or from other changes in network connectivity and dynamics.
  4. The MiniHack tasks involve complex behaviors that may not directly correspond to the simple computational operations defined for isolated FFLs. The paper mainly evaluates overall learning performance, without examining the relevant response dynamics in trained networks.
Questions for authors
  1. What is the main theoretical novelty of the FFL analysis compared with existing work on signed motifs and E-I circuit dynamics? A brief look at Zhang et al, 2026 work, they studied 13 types of connections, which seem already include these signed configs studied in this work.
  2. How does the linearized model explain context-dependent gating, as the effective dynamics may change with the background input and activation state?
  3. Can the authors provide ablation experiments to distinguish the effects of specific signed motifs from other structural properties, such as simple E-I balance or weight distributions?
6 comments

We thank the reviewer for this comment. We compare linear and ReLU FFL dynamics under matched inputs and weights and examine the contributions of direct and indirect pathways. We also analyze event-aligned responses in trained RNNs and perform targeted connection ablations and structural controls. Together, these analyses connect E/I FFL computational biases with task-relevant response dynamics and behavioral performance.


W1 & Q1 — What is the theoretical novelty beyond the shared stability of the four E/I FFLs?

Our theoretical advance is to explain how equally stable E/I FFLs exhibit different computational biases depending on their E/I configurations, and to characterize the conditions supporting these biases. We use signed, weighted summation in a neural network dynamical model. This differs from the promoter-level AND/OR integration studied by Mangan and Alon (2003), requiring us to reexamine how the two pathways combine and shape the response.

Specifically, Proposition 1 separates direct and indirect pathway contributions, helping explain their reinforcement or competition and how their relative influence changes over time. Building on this analysis, comparisons under matched inputs and connection magnitudes reveal distinct tendencies toward response onset, switching, stopping, and context-dependent gating. The input and weight scans further show that these functions occur across a range of input conditions and connection strengths, rather than only at a particular parameter combination. This extends the analysis from shared stability to functional differences within the same stability class.


W2 & Q2 — How does the linearized model explain the four input-dependent computations and context-dependent gating?

Regarding W2, linearization helps explain which functions different combinations of E/I pathways tend to support. Nonlinearity further shapes these functions and determines the conditions under which they emerge. For nondecreasing activations, the choice of activation function or reference equilibrium does not reverse the E/I signs of active pathways. This allows us to use the pathway decomposition in Proposition 1 as a common framework for understanding these computational biases. Specifically, it separates direct and indirect contributions, revealing how they reinforce or compete with each other and how their relative influence changes over time. These pathway interactions help explain the four computations:

  • All-E: Both pathways reinforce the output, supporting earlier response-threshold crossing.
  • FFI: Direct excitation competes with the relatively later influence of indirect inhibition. The direct pathway initially raises the output, while the indirect pathway gradually builds up its inhibitory influence. This competition helps limit the persistence of an old response and facilitate switching.
  • Suppression: The Go input to B promotes the output, while a Stop input to A suppresses it through both pathways, helping stop an ongoing Go response.
  • All-I: When B already provides inhibition, an input to A can inhibit B, thereby reducing B’s inhibition of C and enabling response release.

The added analysis of linear dynamics in the four cases shows that models with linear and nonlinear ReLU activations exhibit qualitatively similar dynamics and functional tendencies under matched inputs and weights (Fig. R1).

Regarding Q2, context-dependent gating requires the same signal to produce response release with background and inhibition without it. The linear-activation model can produce response release, but its net signal effect does not depend on the background: the responses with and without background coincide (Fig. R2). In contrast, ReLU provides the required background selectivity.

With background input, B already inhibits C in both models. A signal to A can inhibit B, weakening B’s inhibition of C and enabling response release. The difference appears without background. A positive signal to A inhibits B, driving B’s state below zero. With linear activation, this negative value becomes B’s output and passes through the negative B→C weight, producing a positive contribution to C. ReLU instead maps B’s negative state to zero output, so the change at B is not transmitted to C. The indirect pathway contributes no disinhibition, leaving A’s direct inhibition of C.

Taken together, linearization helps explain the functional tendencies of different E/I pathway combinations, while nonlinearity further shapes these functions and determines the conditions under which they emerge. The scans in Figures 4 and C1 identify regions of input and parameter space that support these functions, showing that they are not restricted to the representative settings in Figure 3.


W3, W4 & Q3 — Do trained RNNs use the proposed local FFL mechanisms in their task-relevant response dynamics?

The trained RNNs exhibit the predicted response patterns, and removing functionally important FFL-associated connections disrupts both the early Combat response and behavioral success.

Response dynamics in trained RNNs (Fig. R3). We test whether the computational biases identified in isolated FFLs are reflected in the response dynamics of trained RNNs. We compare event-aligned responses between each task-matched motif-enriched RNN and Vanilla across four tasks. The motif-enriched networks exhibit response patterns consistent with the computational biases of their corresponding FFLs: All-E shows a stronger early response in Combat; FFI supports faster response switching when mushrooms become poisonous; Suppression shows response suppression during Stop and recovery after Stop ends in Trap; and All-I shows a larger difference between responses with and without the key.

All-E FFL participation and connection importance (Fig. R4a). We test whether connections participating in more target FFLs contribute more to task behavior. In trained Combat networks, we rank FFL-associated connections by the decrease in opportunity success rate caused by individual connection removal and compare this ranking with the number of All-E FFLs containing each connection. The number of All-E FFLs containing a connection is positively associated with its ablation loss in all ten training seeds, with within-seed Spearman ρ=0.464\rho=0.464–0.6350.635. This association links connections shared by more target FFLs to greater functional importance.

Causal contribution of FFL-associated connections (Fig. R4b–d). We test the causal contribution of these FFL-associated connections to response dynamics and task performance. We compare removal of high-contribution FFL-associated edges with the equal-dose random removal control in Fig. R4b–d. At 1% edge removal, high-contribution removal suppresses the early Combat response and accelerates its decay, whereas random removal largely preserves the response peak (Fig. R4b). This contrast extends to opportunity success rate and episode return across 1%, 2%, and 5% removal, with consistently greater disruption following high-contribution removal (Fig. R4c,d). These results establish a causal contribution of high-contribution FFL-associated connections to both response dynamics and Combat task performance.

These results demonstrate the functional importance of recurrent connections participating in All-E FFLs and link their structural participation to task-relevant response dynamics and behavioral performance.


W3 & Q3 — Can the effects of specific signed motifs be distinguished from E/I balance and weight distributions?

Across these controls, motif-enriched networks retain their overall advantages across the four tasks, through faster learning and/or higher returns within the training budget. We conducted three additional ablations involving initial weight norm, weight distribution, and E/I balance (Fig. R5). Specifically, the comparisons vary the initial Frobenius norm, the fourth moment of nonzero weight magnitudes, or local E/I-balance deviation by 1%, preserving each network’s signed threshold graph. The main qualitative performance patterns therefore persist across the tested initialization variations.

I thank the author for these additional analysis and clarification. The comparisons between linear and ReLU dynamics help clarify the roles of signed pathway interactions and nonlinear activation. One of my concern about the equilibrium of the isolated FFL under constant input has not been answered. Since the paper assumes constant inputs, this equilibrium can be explicitly characterized and analyzed.

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

This work focuses on understanding how small network motifs can affect the ability of recurrent models to perform various tasks. Specifically, the authors focus on the feedforward loop (FFL) motif where a direct path from node A to node C is complemented by an indirect path from A to B to C. While prior work has studied the role of FFLs in artificial and biological networks, this work studies the effect of excitatory-inhibitory (E-I) balance in the context of FFLs. To do this, the authors rely heavily on Zhang et al. 2026 which introduce a mechanism for learning tasks in RNNs while maintaining the frequency of various motifs in the network connectivity. Under this setting, the authors show theoretically and empirically that FFLs augmented with fixed signs on the connections produce distinct computational primitives: early activation, response switching, delayed stopping, and context-dependent gating. In a more complicated MiniHack task, the authors train RNNs with these motifs and show how each computational primitive supports a specific task action (e.g., attack initiation). Furthermore, they show that a policy composed of these pre-trained motif RNNs + a learned gating outperforms a vanilla policy.

Critical strengths

Understanding how small circuits compose to perform complex tasks is timely and important for mechanistic interpretability research. This work, which goes in the opposite direction by building models with interpretable mechanisms built-in, could provide useful insights for the reverse problem.

Critical weaknesses
  • I could not find enough details on the MiniHack experiment to know whether the comparison is fair. In particular, I am concerned that the motif-model has 4x the parameters of the vanilla model, and 4x the total training on each sub-task involved in the composite task.
  • The idea of signed FFLs is not new and has already been theoretically explored (see Magnon and Alon, 2003 which the authors cite), so the first contribution which the authors claim as their own seems to be already published.
  • This work leans heavily on Zhang et al. 2026. The extension to Dale-constrained models is new but only a small change. Thus, I see the primary contribution of this work as a systematic exploration of E-I + FFLs. To this extent, I found the results underwhelming. The toy tasks in Section 3.1 are pretty much hand-engineered to favor one of the four motifs. Similarly, even in the more complex MiniHack task, the four actions each favor a particular motif. While these are interesting results to include, a complete paper should include settings where the appropriate motif is not obvious.
Questions for authors
  1. Can you help address my concerns about the comparison for the MiniHack task? Also, is the vanilla network also Dale-constrained?
  2. Can you clarify which parts of your theory make up your first contribution and how it is distinct from prior work?
  3. Can you extend your evaluation to a task where these motifs are reasonably involved but it isn't so cut and dry that 1 action corresponds to 1 primitive?
3 comments

We thank the reviewer for the comments on theoretical novelty, comparison fairness, and task design. We clarify the differences from prior FFL studies and the shared architecture, training budgets, and Dale constraints of the motif and vanilla systems. We also added an Atari Seaquest experiment, where multiple behavioral demands interact without a prescribed action-to-motif mapping. These comparisons support the benefits of motif enrichment in tasks with interacting demands.


W2 & Q2 — What is new in this work compared with Mangan and Alon (2003) and Zhang et al. (2026)?

Our main contribution is to explain why equally stable E/I FFLs exhibit different computational biases in a neural network dynamical model, and to test whether networks with these biases perform better on corresponding tasks and whether networks with different biases can be coordinated to perform more complex tasks. The specific differences are as follows:

  1. Compared with Mangan and Alon (2003). Their work also studied functional differences among signed FFLs, but used AND/OR-type integration in gene regulation. This differs from the signed, weighted summation commonly used in neural network dynamical models. We therefore formulate and analyze signed FFL dynamics under this neural integration rule to characterize their functional differences.
  2. Compared with Zhang et al. (2026). Their stability result indeed covers all four E/I FFLs. However, this stability classification does not distinguish their computational functions. This is one motivation for studying E/I organization further. We reveal their distinct tendencies toward response onset, switching, stopping, and context-dependent gating, and test their benefits on corresponding tasks. In particular, we also examine whether networks with different motif biases can be coordinated according to context to perform complex tasks involving multiple demands. This coordinated use of distinct functional biases was not examined by Zhang et al. (2026).

W1 & Q1 — Are model capacity, training budgets, and Dale constraints matched between the motif and vanilla systems?

Both conditions use the same four-expert architecture and gating framework. The motif condition combines four motif-enriched experts, whereas the vanilla condition combines four corresponding vanilla experts trained on the same constituent subtasks. Each motif expert is paired with a vanilla expert of the same network size, and the gating networks have the same architecture. Consequently, the two systems have the same total parameter count; the four-expert construction applies to both conditions (Fig. 6a1).

The training budgets are also matched at both stages. During subtask training, each motif expert and its vanilla counterpart receive the same training budget under the same training settings. Both conditions therefore undergo equivalent training across all four subtasks. During composite-task training, the pretrained experts are frozen and the gating networks are trained using the same optimization settings and number of environment interactions. The motif system thus receives neither four times the model capacity nor four times the subtask training of the vanilla system.

Finally, both motif and vanilla networks are subject to Dale’s law. Here, “vanilla” denotes the baseline without the proposed motif enrichment, rather than a network without the Dale constraint. This shared constraint ensures that the comparison evaluates the effect of motif enrichment under matched model capacity, training budgets, and Dale constraints.


W3 & Q3 — Does motif enrichment help on tasks without a predefined one-to-one mapping between actions and computational primitives?

We added an experiment on Seaquest from the Atari Learning Environment, where combat, navigation, diver rescue, and oxygen management interact within a single task. The same action can serve different purposes depending on the situation: moving upward, for example, can help avoid an enemy, approach a diver, or reach the surface to replenish oxygen. Seaquest therefore provides a setting in which successful behavior does not reduce to a predefined one-to-one correspondence between an action and a computational primitive.

As shown in Fig. R6, we compared motif-enriched and vanilla networks over 5 million agent steps. The motif model achieves higher mean training returns through most of the later training period, reaching 812.4 versus 761.6 at the final plotted step—an approximately 6.7% improvement. The curves report the mean rolling return over the latest 100 training episodes per seed, with shaded regions indicating the standard error.

These results support the usefulness of motif enrichment in a task with interacting behavioral demands and no prescribed action-to-motif mapping.

Thank you for taking the time to address my concerns.

W1, W2, Q1, Q2: The response satisfies my concerns. On a second read of the abstract/intro, I do think the authors do a reasonable job of distinguishing from prior work. My only suggestion is that this could be made more explicit for readers less familiar with the background literature. With the additional details provided, I am able evaluate the RL results and believe they fairly support the authors' claims.

W3 & Q3: Thank you for running an additional experiment. Ultimately, I am not sure that motifs enhance the network performance qualitatively nor quantitatively. Perhaps a systematic ablation under this more complex task would benefit the paper's goals, i.e., a more systematic exploration of how E-I motifs affect computation capabilities without an explicit claim of performance improvement.

Right now, I am between raising my score and maintaining it. I will wait to see the other review discussions.

We thank the reviewer for the constructive follow-up. In the revised version, we will incorporate the clarifying statements on prior work from our previous response into the Introduction. We will also strengthen the analysis of how E/I motifs affect computational capabilities by incorporating the response-dynamics and connection-ablation results below. Our MiniHack analyses connect E/I FFL computational biases with task-relevant response dynamics and the functional importance of recurrent connections.

We compare event-aligned responses between motif-enriched and vanilla RNNs across four MiniHack tasks (Fig. R3). The motif-enriched networks exhibit response patterns consistent with their corresponding FFL computational biases: All-E shows a stronger early response in Combat; FFI supports faster response switching when mushrooms become poisonous; Suppression shows response suppression during Stop and recovery after Stop ends in Trap; and All-I shows a larger difference between responses with and without the key.

In Combat, participation in more All-E FFLs is associated with greater connection importance. We count the All-E FFLs containing each FFL-associated connection and measure the decrease in opportunity success rate caused by removing that connection. The two are positively correlated within each of ten training seeds, with Spearman correlations of 0.464–0.635 (Fig. R4a).

Targeted connection ablation disrupts both the early Combat response and behavioral success. We compare removal of high-contribution FFL-associated connections with random removal. At 1% connection removal, high-contribution removal suppresses the early response and accelerates its decay, whereas random removal largely preserves the response peak. High-contribution removal also causes greater disruption of opportunity success rate and episode return across 1%, 2%, and 5% removal (Fig. R4b–d).

These results demonstrate the functional importance of recurrent connections participating in All-E FFLs and link their structural participation to task-relevant response dynamics and behavioral performance.

Reviewer_rS77OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper examines signed feedforward loops, showing how they can support different computations like response initiation, response switching, stopping responses, and context-dependent release. It then introduces regularisation methods for encouraging signed feedforward loops to form within larger networks, and shows that the resulting systems exhibit biases toward one of these four computations in the MiniHack environment. With an added gating mechanism, the system can mix these behaviours to obtain higher returns than the vanilla baseline.

Critical strengths

The paper investigates an interesting architecture and the method for enriching a large network with these circuit motifs works well.

The study connects motifs to certain behavioural capacities, with potential implications for neuroscience.

The paper is clear and easy to follow.

The results and figures are done to a high standard.

Critical weaknesses

The empirical results show some benefit in the combined system, but comparison is only to the unenriched network. It would be strengthened by comparisons to an open benchmark. That said, the work outlines several links to neuroscience rather than machine learning and state of the art performance is not the only goal.

Questions for authors

My interpretation of the supplementary material is that the vanilla networks use the same training hyper parameters as the motif networks. If the hyper parameters are optimised separately for each condition, does the motif advantage still exist? The potential concern would be that the motif initialization might change something nonspecific about the optimal hyper parameters, like the initial weight norm, which can trade off with learning rate.

2 comments

We thank the reviewer for the suggestions on hyperparameter optimization and evaluation beyond MiniHack. We added hyperparameter sweeps for both motif-enriched and vanilla networks, together with controls for initial weight norm, weight distribution, and E/I balance. We also extended the comparison to Seaquest, an Atari benchmark with interacting behavioral demands. These comparisons support the benefits of motif enrichment across optimization settings and beyond MiniHack.


Question: Does the motif advantage persist when hyperparameters are optimized separately for the motif and vanilla networks?

Across the four tasks, these experiments show that the benefits of motif enrichment generally persist under the tested hyperparameter variations, through faster learning and/or higher returns within the training budget. To examine whether the observed advantage depends on the original hyperparameter setting, we conducted additional sweeps for both the motif-enriched and vanilla networks, varying the learning rate, PPO clipping parameter, and entropy coefficient to 0.5×0.5\times, 1×1\times, and 2×2\times their original values (Figs. R7, R8, and R9). Thus, the observed advantage is not restricted to a single shared hyperparameter configuration.

To directly investigate the reviewer’s proposed initialization-related explanation, we also conducted ablations involving initial weight norm, weight distribution, and E/I balance (Fig. R5). Each comparison changes the targeted statistic by 1% while preserving the signed threshold graph. The main qualitative performance advantages persist under these controls. Together, these experiments support the robustness of the motif benefits across the tested optimization settings and structural controls.


Weakness: Do the benefits of motif enrichment extend to an open benchmark beyond MiniHack?

On Seaquest, at the final plotted step, the mean training return is 812.4 for the motif model versus 761.6 for vanilla, corresponding to an approximately 6.7% improvement. We extended our experiments to Seaquest, an Atari environment in the Arcade Learning Environment and an established public benchmark beyond MiniHack. We selected Seaquest because it combines combat, navigation, diver rescue, and oxygen management within a single task. These interacting demands provide a useful setting for testing whether the benefits observed in MiniHack extend beyond our original task suite.

As shown in Fig. R6, we compared motif-enriched and vanilla networks over 5 million agent steps. The motif model achieves higher mean training returns through most of the later training period. Each curve averages the rolling return over the latest 100 training episodes per seed, and the shaded regions indicate the standard error across these seeds.

These results extend our empirical comparison to a public benchmark and provide additional evidence for the benefits of motif enrichment beyond MiniHack.

Thank you for the response, I think these results solidify the soundness further. I will maintain my (already positive) score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Summary of Revisions

We thank all reviewers for their constructive feedback and careful evaluation. We have incorporated the following experiments, analyses, and clarifications into the revised manuscript and appendices.

1. Response dynamics in trained RNNs.
New Section 3.4 and Figure 7 connect the computational biases of isolated FFLs to responses during task execution. The trained networks show stronger early responses, faster switching when mushrooms become poisonous, suppression and recovery around Stop signals, and greater response separation with and without the key.

2. Functional importance of recurrent connections.
Figure 8 shows that All-E FFL participation correlates positively with individual connection-ablation loss in all ten seeds, with Spearman correlations of 0.464–0.635. Targeted edge removal disrupts both response dynamics and behavioral performance more strongly than random removal.

3. Explicit equilibrium and nonlinear dynamics.
We characterize the equilibrium under constant input and explain how it determines effective pathway gains. Matched comparisons between linear and ReLU dynamics clarify the roles of signed pathway interactions and nonlinear activation, including the background selectivity required for context-dependent gating.

4. Robustness across optimization and initialization settings.
We added sweeps of learning rate, PPO clipping, and entropy regularization, together with sensitivity analyses for initial weight norm, weight distribution, and E/I balance. The benefits of motif enrichment generally persist across these comparisons through faster learning, higher returns, or both.

5. Evaluation beyond MiniHack.
We added Atari Seaquest, where multiple behavioral demands interact without a prescribed action-to-motif mapping. Over five million agent steps, the motif model reaches a mean training return of 812.4 versus 761.6 for Vanilla at the final plotted step, an approximately 6.7% improvement.

6. Clearer positioning and experimental details.
The Introduction now clarifies the distinctions from Mangan and Alon (2003) and Zhang et al. (2026), emphasizing computational differences among equally stable E/I FFLs under signed, weighted integration. We also document the shared four-expert architecture, equal parameter counts, and matched training budgets of the motif and Vanilla systems.

Together, these additions link local circuit motifs to response dynamics and behavioral performance. Detailed responses to the reviewers’ comments are provided below.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.