acceptodds
Under review as a conference paper at ICLR 2027

Dissecting DNN Overfitting: The Interplay Between Inference Patterns and Overfitting

Abstract

The overfitting of deep neural networks (DNNs) has rarely been studied from the perspective of their detailed inference patterns. In this paper, we explore whether DNN overfitting can be explained through the generalizability of interactions between input variables encoded by the network. Specifically, we explore this question from four complementary perspectives: (1) We find that overfitted DNNs tend to encode non-generalizable interactions. (2) We discover that removing non-transferable interactions from DNNs effectively mitigates overfitting. (3) This improvement arises because enhancing interaction transferability between two DNNs substantially improves the generalizability of these interactions. (4) We find that among various types of feature consistency across DNNs, improving interaction transferability is the most effective in mitigating overfitting.

Primary area:
learning theory
Keywords:
Representation Analysis, Generalization Theory, Neural symbolic

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pShEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a regularization term for network training that helps deal with overfitting. The regularization term couples two networks trained together from different initializations, and penalizes the difference of the two networks. By "difference" it measures how one network's output is affected by a particular combination of input patches. Namely they measure the change in the true-class log-odds that are caused by co-appearence of patch groups and compare this effect of one network with the other.

The experiments train VGG, ResNet and AlexNet on CIFAR-10 and Tiny-ImageNet, BERT on SST-2, and fine-tune Qwen2.5-3B with LoRA. With the regularization term the gap between test and training loss becomes smaller.

Critical strengths

1 Employing cross-network agreement on second-order differences as a training loss is a further development within the line of research on game-theoretic interactions in DNNs.

2 The direction of the effect is consistent across CNNs, BERT and a 3B LLM.

Critical weaknesses

(1) Novelty is marginal. The essential novel idea boils down to an additional training regularization.

While a significant part of the paper is devoted to the technical stack: interaction computation, the sparsifying AND-OR split, the patch selection and the reference-network test of generalizability. These tools/skills are taken from earlier works. Findings from a line of research, e.g. Ren et al. (2024b) and He et al. (2025), are repeated.

This also includes Zhang et al. (2024) and Cheng et al. (2025), which are not cited.

The loss itself is close in form to Zhang et al. (2021), which is not cited either.

References

  • Ren, Q., Zhang, J., Xu, Y., Xin, Y., Liu, D., Zhang, Q. (2024b). "Towards the dynamics of a DNN learning symbolic interactions." NeurIPS 2024. arXiv:2407.19198.
  • He, Y., Zhang, J., Cheng, L., Zhang, H., Zhang, Q. (2025). "Technical report: Quantifying and analyzing the generalization power of a DNN." arXiv:2505.06993.
  • Zhang, J., Li, Q., Lin, L., Zhang, Q. (2024). "Two-phase dynamics of interactions explains the starting point of a DNN learning over-fitted features." arXiv:2405.10262.
  • Cheng, L., Zhang, J., Ren, Q., Zhang, Q. (2025). "Revisiting generalization power of a DNN in terms of symbolic interactions." arXiv:2502.10162.
  • Zhang, H., Li, S., Ma, Y., Li, M., Xie, Y., Zhang, Q. (2021). "Interpreting and boosting dropout from a game-theoretic view." ICLR 2021. arXiv:2009.11729.

(2) The motivation of the first finding is not clearly stated. The paper does not say whether the samples on which interactions are analysed are training or test samples. The conclusion even suggests they are training samples ("pinpoint the specific set of training samples on which such overfitted interactions emerge", Sec. 4). If so, an interaction that the target network encodes on its own training sample, but a reference network that never saw this sample does not encode, is just what overfitting means. Then "overfitted networks encode more non-generalizable interactions" restates overfitting rather than explains it.

Questions for authors

What does it mean to "explain" overfitting with non-generalizable interactions? An interaction is called non-generalizable when the target network encodes it on a sample but a reference network that never saw the training data does not.

9 comments
AuthorsOpenReview ↗

Response to Reviewer pShE (Part 1/2)

Thank you for your great efforts on the review. We will try our best to answer all your questions.

If any answer is unclear or you have further questions, please let us know so we can respond before the deadline.


Q1: Ask about the relation to prior interaction-based work

(1) Novelty is marginal. The essential novel idea ... cited either.

A: A good question. This paper advances the research frontier of interaction-based explanation from the conventional topics of “interaction extraction” and “phenomenological analysis” to interaction-rectification-based generalizability enhancement.

Interactions offer a new mathematical representation tool for characterizing generalization of a DNN, and this tool has already been widely used [cite1, cite2, cite3]. However, previous studies are mainly limited to the scope of explaining and analyzing DNN generalization, e.g., examining DNN training dynamics [cite4, cite6], and measuring interaction generalizability [cite5, cite7].

By contrast, our work is the first, to the best of our knowledge, to move the research frontier from interaction extraction and analysis to interaction-rectification-based generalizability enhancement. Specifically, we explore a new methodology for boosting a model’s generalizability, i.e., directly penalizing non-generalizable interactions. Experiments show that our interaction-based training method reduces the model's loss gap much more significantly than standard regularization methods.

In sum, this work, for the first time to the best of our knowledge, advances the frontier of interaction-based explanation to the scope of interaction-rectification-based generalizability enhancement, and significantly reduces model overfitting.

In addition, regarding the paper [cite8] that you highlight, [cite8] shows that dropout mainly penalizes high-order interactions, but this paper does not rigorously disentangle non-generalizable interactions. Therefore, merely optimizing interaction complexity cannot improve performance as effectively as our method. Nevertheless, we will carefully cite and discuss all the above papers.

[cite1] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite2] Explaining Generalization Power of a DNN using Interactive Concepts. AAAI, 2024.

[cite3] Does a Neural Network Really Encode Symbolic Concept? ICML, 2023.

[cite4] Towards the dynamics of a DNN learning symbolic interactions. NeurIPS, 2024.

[cite5] Technical report: Quantifying and analyzing the generalization power of a DNN. arXiv:2505.06993.

[cite6] Two-phase dynamics of interactions explains the starting point of a DNN learning over-fitted features. arXiv:2405.10262.

[cite7] Revisiting generalization power of a DNN in terms of symbolic interactions. arXiv:2502.10162.

[cite8] Interpreting and boosting dropout from a game-theoretic view. ICLR, 2021.


AuthorsOpenReview ↗

Response to Reviewer pShE (Part 2/2)

Q2: Ask which samples are used to evaluate interaction generalizability and whether the finding merely restates overfitting

(2) The motivation of the first finding is not clearly stated ... restates overfitting rather than explains it.

A: A good question. We have several perspectives from which to address your concern.

First, let us clarify the difference in the experimental settings between Section 2.2 and Section 2.3. In Section 2.3, two target DNNs are trained on the same data to measure interaction transferability. Therefore, the interactions are evaluated on the training set, and the interaction patterns in Figure 3 represent patterns that can be transferred across different DNNs. Such patterns are generally more reliable. In Section 2.2, the baseline DNN is trained on data that are unseen by the target DNN. The samples used to compute the interaction distributions in Figure 2 are test samples of the target DNN. Therefore, the interaction patterns in the target DNN, which can be transferred to baseline DNNs, actually represent patterns generalizable to unseen data (training samples of the baseline DNN are not used to train the target DNN). If an interaction pattern can be transferred to the baseline DNNs, this interaction pattern can also serve as a common inference pattern for the unseen data and is therefore generalizable [cite1, cite2].

Second, whether interaction generalizability is a restatement of model generalization or an explanation of model generalization is an insightful point. To some extent, yes: our study provides a fine-grained characterization of the overfitting phenomenon with validated practical value in model training, rather than an explanation of the fundamental causes of overfitting.

However, the exact causes of overfitting are naturally ambiguous and lack strict analytic formulations. In comparison, we adopt a reductionist perspective: a precise and structured mechanistic description of the overfitting phenomenon itself would contribute more to substantive progress in the field than qualitative findings that lack strict constraints.

Historically, because overfitting has not been analyzed in sufficient detail, practitioners have often had to rely on various regularization techniques to address it, with limited efficiency. As shown in Figure 4 and Table 2, once we can accurately identify and intervene on non-generalizable interactions, the loss gap decreases much more efficiently in our experiments.

[cite1] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite2] Does a Neural Network Really Encode Symbolic Concept? ICML, 2023.


Q3: Ask what it means to explain overfitting with non-generalizable interactions

What does it mean to "explain" overfitting ... does not.

A: Thank you for pointing out the imprecise wording. We will describe our contribution more precisely as a fine-grained characterization and intervention of overfitting, thereby removing the ambiguity that the word “explain” may have introduced in our previous wording.

Our work does not attempt to explain the fundamental features, root causes, or training settings that lead to overfitting. Instead, we use interactions as a new mathematical tool to reveal local details of the global overfitting phenomenon. Overfitting, traditionally viewed as a property of the model as a whole, is represented in terms of the aggregate generalizability of the interaction patterns encoded by the model.

More crucially, this decomposition of overfitting provides precise guidance for further mitigating DNN overfitting. Our experiments in Figure 4 and Table 2 also show that penalizing non-generalizable interactions reduces the loss gap in DNNs more efficiently than competing regularization methods.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewer pShE,

Thank you again for your comments and for taking the time to review our responses. As the author–reviewer discussion period will end on November 18, we would sincerely appreciate any feedback on whether our responses have addressed your concerns.

If any concerns remain, please let us know, and we would be happy to provide further clarification or additional analyses before the discussion period closes.

Many thanks for your time and consideration.

Reviewer_2PuQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Assuming that an input x composed of N variables, [x_1, x_2, ... X_N], their AND and OR inference patterns are introduced. With inference patterns, overfitting of deep neural networks (DNNs) are evaluated. Specifically, empirically overfitting is explored from four perspectives: (1) it is shown that overfitted DNNs tend to encode "non-generalizable interactions"; (2) it is shown that reducing "non-transferable interactions" from DNNs mitigates overfitting; (3) It is shown that enhancing interaction transferability between two DNNs substantially improves the "generalizability of these interactions"; (4) from empirical evaluation it is shown that improving "interaction transferability" is the most effective in mitigating overfitting.

Critical strengths

S1. The paper attempts to empirically establish that if an input is composed of N separable variables, then ANN and OR functions of some or all subsets of these variables can be used to characterize the overfitting of DNNs. S2. The reported work uses this concept to connect "non-generalizable interactions". S3. The reported work uses this concept to connect "non-transferable interactions". S4. The paper shown how simultaneous retaining of two (or more) overfitted networks can be utilized to reduce their overfitting.

Critical weaknesses

W1. The term generalizability and transferability are quite well-established concepts. But the use of these terms without redlining them is very confusing and it will be same for all readers, I strongly believe. W2. The paper is not stand alone. One has to read the appendices to follow what the work is about. W3. If an input x is decomposed into N variables, then there are 2^N possible subset and for each of them AND and OR combinations has to be considered for completeness. This will produce too many possible combinations of these variables. For example, consider ViT: a 224x224 image will create 784 patches and if each patch is considered as one variable, then 2^784 combinations ought to be coded and evaluated. An impossible task. W4. It is concluded that "overfitted DNNs" tend to encode "non-generalizable interactions". But as discussed above, for practical DNNs the methos is computationally impractical, actually impossible. W5. Similar and much bigger problems will appear in other three studies reported. W6. The paper also does not compare standard regularization methods with their methods. W7. Overfitting problem is well known and there are many established practical methods to mitigate it. For example, partitioning of available examples into training, validation, and evaluation sets. W8. Overall, the paper contributed very little either practically or theoretically to the understanding or mitigating overfitting problem.

Questions for authors

Q1. Compare computing time ratio of the proposed method to the training-validation-evaluation method. Use available data regularization method with training-validation-evaluation method. Can you do this with CIFAR-10 dataset and ResNet50? Q2. If one uses training-validation-evaluation method and the proposed method, which one will have higher accuracy on CIFAR-100 dataset with ResNet50 or a Densenet121? For training-validation-evaluation method use available regularization method.

10 comments
AuthorsOpenReview ↗

Response to Reviewer 2PuQ (Part 1/6)

Thank you for your great efforts on the review. We will try our best to answer all your questions.

If any answer is unclear or you have further questions, please let us know so we can respond before the deadline.


Q1: Ask about the definitions of generalizability and transferability

W1. The term generalizability and transferability are ... I strongly believe.

A: Thank you. For interaction patterns, generalizability and transferability are closely related but distinct concepts. We define and explain interaction generalizability and interaction transferability separately in Sections 2.2 and 2.3, respectively.

Interaction generalizability indicates whether an interaction pattern can be transferred to other DNNs trained on unseen data (Line 159), whereas interaction transferability indicates whether the same interaction pattern is consistently encoded when two different DNNs make predictions on the same input sample (Line 244). Specifically, the two DNNs used for evaluating interaction transferability are trained on the same training set.

However, unlike the measurement of interaction transferability, directly measuring interaction generalizability requires exhaustively examining the occurrence of each specific interaction over a large number of unseen samples, which would incur prohibitive computational costs. Thus, we adopt an efficient approximation strategy proposed by [cite1, cite2], which quantifies an interaction’s generalizability through its transferability to baseline DNNs trained on different training data. This is because, from the perspective of the target DNN, the training data used by the baseline DNNs are unseen. If an interaction pattern can be transferred to the baseline DNNs, this interaction pattern can also serve as a common inference pattern for the unseen data. In this way, transferability to DNNs trained on unseen data has been considered a faithful indirect approach for measuring interaction generalizability [cite1, cite2, cite3].

We will further clarify the definitions of these two terms and the distinction between them in the revised manuscript to avoid confusion.

[cite1] Technical report: Quantifying and analyzing the generalization power of a dnn. arXiv preprint arXiv:2505.06993, 2025.

[cite2] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite3] Does a Neural Network Really Encode Symbolic Concept? ICML, 2023.


AuthorsOpenReview ↗

Response to Reviewer 2PuQ (Part 2/6)

Q2: Ask for more technical details in the main paper

W2. The paper is not stand alone. One has to read the appendices to follow what the work is about.

A: We fully understand the concern of the reviewer. Indeed, most advanced machine learning theories, such as generalization theory [cite1, cite2], optimization theory [cite3, cite4], and neural tangent kernel theory [cite5], have their own technical foundations and specialized notation. It is therefore natural that readers may need some time to become familiar with them. This is an unavoidable cost of reviewing advanced theoretical work, but we do not believe that technical complexity alone is equivalent to an incomplete technical presentation.

More importantly, our paper provides a level of technical completeness comparable to previous ICLR/ICML/NeurIPS papers on interaction-based explanation [cite6, cite7, cite8]. All necessary information for understanding interactions has already been introduced in Section 2.1.

Building on this established theory, the main paper's “2.1 Preliminaries: Interactions” section introduces the basic concepts, including the definition of interactions, the AND and OR logical model, interaction sparsity, the universal matching property, and the basic concept of interaction generalizability. In addition, implementation details, including how interactions are specifically extracted, how the input variables are selected, and how the interaction-based loss is implemented, have already been introduced in Appendices G, F.5, and B, respectively, although these implementation details do not directly affect the presentation of the core ideas of this paper.

Nevertheless, we will further polish the presentation of the relevant technical details in the revised manuscript and provide a clearer technical framework.

Despite the technical complexity of interaction-based explanation, this perspective offers a fine-grained view to revisit and decompose a DNN’s generalization capacity into compositional inference patterns, and, for the first time, enables us to mitigate DNN overfitting through interventions at the level of compositional inference patterns.

[cite1] Understanding Deep Learning Requires Rethinking Generalization. ICLR, 2017.

[cite2] A Unified Framework for Information-Theoretic Generalization Bounds. NeurIPS, 2023.

[cite3] On the Convergence of Adam and Beyond. ICLR, 2018.

[cite4] On Convergence of Adam for Stochastic Optimization under Relaxed Assumptions. NeurIPS, 2024.

[cite5] Neural Tangent Kernel: Convergence and Generalization in Neural Networks. NeurIPS, 2018.

[cite6] Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs. ICLR, 2024

[cite7] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite8] Explaining Generalization Power of a DNN using Interactive Concepts. AAAI, 2024.


AuthorsOpenReview ↗

Response to Reviewer 2PuQ (Part 3/6)

Q3: Ask about the scalability and practical feasibility of interaction computation

W3. If an input x is decomposed into N variables ... An impossible task.

W4. It is concluded that "overfitted DNNs" tend ... computationally impractical, actually impossible.

W5. Similar and much bigger problems will appear in other three studies reported.

A: Thank you. Please see the “Discussion on Computational Complexity” paragraph in Section 2.1, Appendix F.5, and Appendix H. We would like to clarify that such theoretically exponential computational costs do not necessarily lead to unaffordable costs in real-world applications, nor do they inevitably prevent practical applications. In fact, a range of approximation algorithms and engineering techniques [cite1, cite2, cite3] have been developed to substantially reduce the cost of interaction computation, and interaction-based explanation has already been applied to the analysis of various large-scale DNNs [cite4, cite5].

Reducing the cost of interaction extraction. Previous work has proposed several efficient methods for extracting interactions. For example, the Sparse Möbius Transform [cite1] exploits interaction sparsity to substantially reduce the exponential cost of exact interaction extraction. Spectral Explainer [cite2] efficiently identifies interactions in LLMs with long inputs (n=1000n=1000), while ProxySPEX [cite3] improves scalability via lightweight proxy models. These methods and related engineering techniques are discussed in detail in Appendix H.

Furthermore, practical interaction analysis does not require treating every raw input element as an input variable. Previous studies [cite4, cite6] have found that DNN predictions usually depend mainly on a small number of salient image regions or meaningful text segments. Accordingly, classical settings for interaction analysis typically select about 10 to 12 input regions or text segments (e.g., words, phrases, sentences, or even sections) for analysis, without enumerating combinations of all pixels or patches. More crucially, the above setting of input variables does not affect the effectiveness of interaction generalizability analysis in previous studies [cite5, cite6]. Thus, we follow this common setting in practice. Appendix F.5 provides the implementation details.

Linear cost of the interaction reduction loss. More importantly, the interaction reduction loss has a fast approximate optimization algorithm in Appendix B, which does not require explicitly extracting or enumerating all interactions in advance. As shown in Appendix B, the training loss we propose has essentially linear time complexity.

Therefore, the proposed interaction optimization method in this paper is computationally feasible.

[cite1] Learning to Understand: Identifying Interactions via the Mobius Transform. NeurIPS, 2024.

[cite2] SPEX: Scaling Feature Interaction Explanations for LLMs. ICML, 2025.

[cite3] ProxySPEX: Inference-Efficient Interpretability via Sparse Feature Interactions in LLMs. NeurIPS, 2025.

[cite4] Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs. ICLR, 2024

[cite5] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite6] Does a Neural Network Really Encode Symbolic Concept? ICML, 2023.


Reviewer_ZVSWOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper explains DNN overfitting through interactions, which are groups of input variables whose joint presence changes the model output. It shows that overfitted models encode more interactions that do not appear in models trained on other data, and that a loss making two models trained together encode the same interactions reduces the gap between training and testing loss. Experiments on CNNs, BERT models and Qwen2.5-3B also show that the shared interactions generalize better to other models, and that output or feature consistency losses do not have the same effect.

Critical strengths

Strengths

  1. The paper explains overfitting through the input combinations that a model actually uses, which gives a view at the level of individual inference patterns. Most prior analyses, such as generalization bounds and loss landscape flatness, treat generalization as a property of the whole model, as Section 3 notes. Here one can see which interactions appear when the training and testing losses start to separate.

  2. The paper turns an analysis tool into a training method that does not need test data. Earlier definitions of generalizable interactions compare against a model trained on test samples. Section 2.3 instead trains two models on the same data and uses each as the reference for the other, and Appendix B gives an approximate loss that avoids computing all interactions.

  3. The experiments go from observation to intervention. Section 2.2 shows that interactions which do not generalize grow after overfitting starts, Section 2.3 shows that removing them shrinks the loss gap, and Figure 13 shows that the gap widens again when the loss is switched off and shrinks when it is switched back on. This gives stronger support for the link than a correlation alone.

  4. The findings hold across a wide range of models and data. The experiments cover AlexNet, VGG, ResNet and BERT models on image and text classification, and Appendix E.1 repeats the main result on Qwen2.5-3B fine tuned with LoRA, where the loss gap drops from 0.354 to 0.034.

  5. The main results are stable across random seeds and hyperparameter settings. Tables 3 to 5 report means and standard deviations over four seeds, and Figures 3 and 11 show similar gains in interaction transferability for λ = 0.3 and λ = 0.8.

Critical weaknesses

Weaknesses

  1. (Solved) The relation to closely related prior work is not well discussed, so it is hard to tell which findings are new. He et al. [1] already report that the overfitting phase is dominated by complex interactions that do not generalize, which matches the first finding, but the paper cites this work only for the metric. Penalizing interaction strength to reduce overfitting was proposed by Zhang et al. [2], and training two networks to match each other was proposed in Deep Mutual Learning [3].

  2. (Solved) The smaller loss gap may not come only from removing interactions. In Equation 7 the cross entropy weight is 1 − λ, so a larger λ makes the model fit the training data less, which by itself raises the training loss and narrows the gap in Figure 4(b). Also, the approximate loss in Appendix B is computed on many masked inputs, while the competing losses L_z and L_f use only the original inputs. The gain over them may therefore come from consistency on masked inputs rather than from interactions.

references:

[1] Yuxuan He, Junpeng Zhang, Lei Cheng, Hongyuan Zhang, and Quanshi Zhang. Technical report: Quantifying and analyzing the generalization power of a dnn. arXiv preprint arXiv:2505.06993, 2025.

[2] Hao Zhang, Sen Li, Yinchao Ma, Mingjie Li, Yichen Xie, and Quanshi Zhang. Interpreting and boosting dropout from a game-theoretic view. In International Conference on Learning Representations, 2021.

[3] Ying Zhang, Tao Xiang, Timothy M Hospedales, and Huchuan Lu. Deep mutual learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4320–4328, 2018.

Questions for authors

If L_z is computed on the same masked inputs used by the approximate loss in Appendix B, does it still fail to reduce the loss gap?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ZVSW (Part 1/2)

Thank you for your great efforts on the review. We will try our best to answer all your questions.

If any answer is unclear or you have further questions, please let us know so we can respond before the deadline.


Q1: Ask for more discussion of closely related work

The relation to closely related prior work ... Deep Mutual Learning [3].

A: A good question. This paper advances the research frontier of interaction-based explanation from the conventional topics of “interaction extraction” and “phenomenological analysis” to interaction-rectification-based generalizability enhancement.

Interactions offer a new mathematical representation tool for providing a fine-grained characterization of DNN generalization, and this tool has already been widely used [cite4, cite5, cite6]. However, previous studies are mainly limited to the scope of explaining and analyzing DNN generalization, e.g., measuring interaction generalizability [cite1, cite5, cite6].

By contrast, our work is the first, to the best of our knowledge, to move the research frontier from interaction extraction and analysis to interaction-rectification-based generalizability enhancement. Specifically, we explore a new methodology for boosting a model’s generalizability, i.e., directly penalizing non-generalizable interactions. Experiments show that our interaction-based training method reduces the model's loss gap much more significantly than standard regularization methods.

In sum, this work, for the first time to the best of our knowledge, advances the frontier of interaction-based explanation to the scope of interaction-rectification-based generalizability enhancement, and significantly reduces model overfitting.

In addition, the two papers you mention [cite2, cite3] are two classic works, but they focus on problems different from ours. First, [cite2] shows that dropout mainly penalizes high-order interactions, but it does not characterize overfitting in terms of non-generalizable interactions. Therefore, merely optimizing interaction complexity cannot improve performance as effectively. Second, [cite3] proposes a mutual-distillation algorithm to co-train two networks, but it does not provide a fine-grained characterization of overfitting at the interaction level. We believe that precisely extracting non-generalizable representations is the key reason why our interaction-penalization algorithm outperforms other regularization methods (e.g., those in Table 2). Nevertheless, we will carefully cite and discuss these works.

Besides, the referenced work [cite1] is a four-page technical report, and it does not discuss interaction optimization at all. It is not a formal paper and does not constitute a direct competing submission. According to the ICLR review policy, authors are not required to compare against arXiv papers, let alone a four-page technical report that does not discuss interaction optimization.

[cite1] Technical report: Quantifying and analyzing the generalization power of a dnn. arXiv preprint arXiv:2505.06993, 2025.

[cite2] Interpreting and boosting dropout from a game-theoretic view. ICLR, 2021.

[cite3] Deep mutual learning. CVPR, 2018.

[cite4] Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in DNNs. ICLR, 2024

[cite5] Defining and Extracting Generalizable Interaction Primitives from DNNs. ICLR, 2024.

[cite6] Explaining Generalization Power of a DNN using Interactive Concepts. AAAI, 2024.


AuthorsOpenReview ↗

Response to Reviewer ZVSW (Part 2/2)

Q2: Ask about possible confounding factors in the loss-gap reduction

The smaller loss gap may not come only from removing interactions ... masked inputs rather than from interactions.

If LzL_z is computed on the same masked inputs ... reduce the loss gap?

A: A good question. The first issue is whether the reduction in the loss gap results from (1) Case 1, “the new loss simply taking weight away from the cross-entropy objective,” or (2) Case 2, “the new loss function really helps prevent overfitting.” The core difference between these two cases can be tested directly: Case 1 will not help reduce the test loss, whereas Case 2 will directly reduce the test loss.

Experimental results in Figure 4 show that our proposed loss significantly reduces the test loss. This successfully demonstrates that the new loss belongs to Case 2, i.e., the loss itself really helps mitigate DNN overfitting.

The second issue is whether the advantage of our method over the competing methods LzL_z and LfL_f comes solely from using masked inputs. To investigate this, we follow the reviewer's suggestion and conduct a new ablation experiment. Specifically, we compute the competing losses LzL_z and LfL_f using exactly the same masked inputs as those used in Appendix B:

Lzmask=Ex[λET∈M∥z(xT)−z′(xT)∥22+(1−λ)(LCE(v)+LCE(v′))],\mathcal{L}_{z}^{\mathrm{mask}} = \mathbb{E}_{x} \left[ \lambda \mathbb{E}_{T\in\mathcal{M}} \left\|z(x_T)-z'(x_T)\right\|_2^2 + (1-\lambda) \left( \mathcal{L}_{\mathrm{CE}}(v) + \mathcal{L}_{\mathrm{CE}}(v') \right) \right],

Lfmask=Ex[λET∈M∥f(xT)−f′(xT)∥22+(1−λ)(LCE(v)+LCE(v′))],\mathcal{L}_{f}^{\mathrm{mask}} = \mathbb{E}_{x} \left[ \lambda \mathbb{E}_{T\in\mathcal{M}} \left\|f(x_T)-f'(x_T)\right\|_2^2 + (1-\lambda) \left( \mathcal{L}_{\mathrm{CE}}(v) + \mathcal{L}_{\mathrm{CE}}(v') \right) \right],

where xTx_T denotes a masked input sample, and M\mathcal{M} denotes the same set of masked states as in our work. In this way, all methods use the same masked samples for training, ensuring a fair comparison. We conduct experiments with VGG-11, VGG-13, ResNet-20, and ResNet-32 on the CIFAR-10 dataset.

The following table compares the training-test loss gap obtained by different methods using the same masked inputs. Lower values indicate smaller loss gaps.

MethodVGG-11VGG-13ResNet-20ResNet-32
Normal training1.07230.76940.95571.0642
Competing method Lfmask\mathcal{L}_{f}^{\mathrm{mask}}1.32851.09441.25511.3884
Competing method Lzmask\mathcal{L}_{z}^{\mathrm{mask}}0.89260.73840.69320.7083
Our method0.24260.19020.18320.2097

The above table shows that our method is much more effective than the two competing methods LzL_z and LfL_f at reducing the loss gap. This means that the advantage of our method does not simply come from using masked inputs; it mainly comes from eliminating interactions that do not generalize.

Dear authors,

Thanks for the rebuttal, all my concerns have been resolved.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and Area Chair

We sincerely thank all reviewers and the Area Chair for their time and efforts in evaluating our paper. We appreciate the constructive comments and have provided detailed responses to all concerns. We would like to briefly summarize the key contributions of our work and the additional evidence provided during the rebuttal.

1. Main contributions of our work. Our work investigates DNN overfitting from the perspective of the generalizability of compositional inference patterns. Specifically, we show that overfitted DNNs tend to encode more non-generalizable interactions, and that selectively suppressing such interactions can effectively mitigate overfitting. More importantly, our work moves interaction-based explanation beyond the analysis of inference patterns toward direct intervention and optimization of their generalizability. This provides a new form of interaction-level supervision that complements conventional output-level supervision.

2. Additional experiments supporting our findings. During the rebuttal, we conducted additional experiments to directly address the reviewers' concerns and examine possible alternative explanations for our results. First, experiments on ResNet-50 and DenseNet-121 using CIFAR-100 demonstrate that our method achieves smaller generalization gaps and higher test accuracy than conventional L1 and L2 regularization. Second, additional ablation experiments using identical masked inputs show that our method substantially outperforms output-consistency and feature-consistency objectives. These results provide further evidence that the effectiveness of our method cannot be attributed solely to the use of masked inputs, reduced emphasis on the cross-entropy objective, or conventional cross-model consistency.

3. Clarifications regarding the technical contributions and feasibility. We have further clarified the distinctions between our work and closely related studies, e.g., Zhang et al. (ICLR 2021) and Deep Mutual Learning, in terms of the interaction formulation, optimization objective, and training mechanism. We have also clarified the computational feasibility of our approach, particularly the use of an efficient approximation of the interaction-based training loss. Furthermore, we have clarified that our explanation of overfitting focuses on how overfitting manifests through the internal inference patterns encoded by DNNs, rather than on the external causal factors that lead to overfitting.

In summary, we believe that our work makes a meaningful contribution by connecting the generalizability of compositional inference patterns to DNN overfitting and demonstrating that directly optimizing these patterns provides an effective approach to mitigating overfitting. The additional experiments and clarifications provided during the rebuttal further strengthen our findings and address the major concerns raised during the review and discussion.

We sincerely thank the reviewers and Area Chair again for their feedback and consideration. We hope that these additional results and clarifications will be taken into consideration in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.