Efficient Activation Function Optimization through Surrogate Modeling

Carefully designed activation functions can improve the performance of neural networks in many machine learning tasks. However, it is difficult for humans to construct optimal activation functions, and current activation function search algorithms are prohibitively expensive. This paper aims to improve the state of the art through three steps: First, the benchmark datasets Act-Bench-CNN, Act-Bench-ResNet, and Act-Bench-ViT were created by training convolutional, residual, and vision transformer architectures from scratch with 2,913 systematically generated activation functions. Second, a characterization of the benchmark space was developed, leading to a new surrogate-based method for optimization. More specifically, the spectrum of the Fisher information matrix associated with the model's predictive distribution at initialization and the activation function's output distribution were found to be highly predictive of performance. Third, the surrogate was used to discover improved activation functions in several real-world tasks, with a surprising finding: a sigmoidal design that outperformed all other activation functions was discovered, challenging the status quo of always using rectifier nonlinearities in deep learning. Each of these steps is a contribution in its own right; together they serve as a practical and theoretical foundation for further research on activation function optimization.

Paper

Similar papers

Peer review

Reviewer sps38/10 · confidence 4/52023-07-02

Summary

The paper presents a new method for improving the performance of neural networks through the design of optimal activation functions. The authors created benchmark datasets by training convolutional, residual, and vision transformer architectures with systematically generated activation functions. They then developed a new surrogate-based method for optimization, which uses the spectrum of the Fisher information matrix and the activation function's output distribution to predict performance. The method was tested on CIFAR-100 and ImageNet tasks, and the results showed significant improvements in accuracy.

Strengths

1. This paper introduces an innovative approach to enhancing activation functions, surpassing existing techniques in both efficiency and effectiveness. 2. The paper is exceptionally well-written, and the experiments conducted are notably thorough. 3. The benchmark datasets created by the authors provide a foundation for future research on activation function properties and their impact on performance.

Weaknesses

None

Questions

None

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Yes

Reviewer x7yQ7/10 · confidence 4/52023-07-06

Summary

This paper introduces three benchmark datasets created by training CNN, ResNet, and ViT architectures using a set of activation functions generated from a three-node computation graph that combines unary and binary operations. The benchmarks serve to showcase the efficacy of utilizing the 2D UMAP of the Fisher information matrix (FIM) spectrum and/or activation outputs as a cost-effective surrogate for predicting activation performance. Leveraging the 2D feature representation, an efficient activation optimization method, AQuaSurF, is developed by employing regression techniques to model activation accuracy across the 2D feature space, requiring only 100 function evaluations. The benchmark results further demonstrate the effectiveness and statistical reliability of this approach. The proposed method is successfully applied to various vision tasks, where the discovered activation functions consistently outperform existing baseline activations. Moreover, the top activations identified through this search exhibit successful transferability to a new vision task.

Strengths

The paper is well-written and easy to follow. The approach of utilizing the UMAP embedding of the FIM spectrum with activation outputs to assess activation performance is novel and interesting. In contrast to previous methodologies that relied mostly on evolutionary algorithms and required thousands of function evaluations, the method proposed in this work demonstrates efficiency by outperforming baselines with just 100 function evaluations. Furthermore, the benchmark datasets introduced in this work, may potentially help accelerate research on activation optimization. Overall, this paper offers valuable insights for assessing activation performance and also introduces a more efficient methodology for activation optimization.

Weaknesses

In Section 6, the authors apply their proposed method to more challenging datasets and a larger activation search space, compared to those used to create the benchmarks. To further evaluate the effectiveness of the approach it would be beneficial to apply the method (KNR on UMAP embeddings) to vision tasks involving new network architectures as well. While the chosen baseline activations in Table.1 already include ReLU and Swish, used in the original three architectures studied in the paper, in order to further strengthen the results it would still be advantageous, and perhaps straightforward, to extend the list of baselines at least to those used in PANGAEA, including GELU, LeakyReLU etc.

Questions

1- In the first paragraph of page 3 the authors observe, based on the scatter plots in Fig 2, that "best results come from discovering functions specialized to individual tasks". However, upon comparing the upper-left and lower-right corners of the plots with the upper-right region it appears that the best functions on one task also transfer effectively to, and are potentially among the best on, the other task. Is this interpretation correct? 2- In the middle row of Fig 4. The UMAP depends only on the activations and not the model. However, there appears to be differences in the distribution of points in the 3 plots (and also compared to Fig.3). Is this because of filtering out failed activations and possible rescalings / reflections of the space? A brief comment on this would enhance clarity for readers. 3- On lines 228-229 of the manuscript "Thus, activation functions are embedded close to each other in this space if they have similar shapes, if they induce similar FIM eigenvalues, or both", considering that the metric on the union of the representations is the sum of the metrics on the individual representations, then shouldn't the activations be close to each other only if they have both similar shapes and similar FIM eigenvalues? 4- How does AQuaSurF compare with PANGAEA in terms of performance? given the partial similarity of the search spaces, is it possible to make a direct comparison between the two methods (e.g. by limiting the space to non-parametric functions)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are partly addressed in the Future Work section in the appendix. There are no concerns regarding negative societal impact.

Reviewer UeiT3/10 · confidence 5/52023-07-06

Summary

This paper addresses the optimization of activation functions in neural networks for improved performance in machine learning tasks. The authors create benchmark datasets and propose a surrogate-based optimization method based on a characterization of the benchmark space. They apply this method to discover better activation functions in CIFAR-100 and ImageNet tasks, showcasing its practical effectiveness.

Strengths

1. The authors create benchmark datasets (Act-Bench-CNN, Act-Bench-ResNet, and Act-Bench-ViT) by training various architectures with numerous activation functions. 2. The paper presents a novel surrogate-based optimization method that characterizes activation functions analytically. By utilizing the Fisher information matrix's spectrum and activation function output distribution, a low-dimensional representation is created. 3. The proposed method, AQuaSurF, efficiently discovers improved activation functions in CIFAR-100 and ImageNet tasks, surpassing previous approaches in terms of evaluation efficiency.

Weaknesses

1. The motivation and definition of using "Activation Function Outputs" as feature in Section 3 is not clearly explained. 2. In Table 1, some widely used human-designed activation functions, such as ELU, ReLU, and Swish, consistently achieve top performance on various tasks with different networks. However, the top activation functions discovered by the proposed method vary across tasks and networks. This suggests a limited generalizability of the searched activation functions. In other words, when faced with a new task or utilizing a new network, the activation function needs to be searched again. Furthermore, this also implies that the searching method may overfit the specific task and network, rather than finding activation functions that are generally effective and meaningful. 3. Related to the previous point, the design of the search space appears overly complicated, which also raises concerns about overfitting. As observed, the top activation functions discovered through the search process often involve complex combinations of existing human-designed activation functions. This complexity reduces their interpretability. Human-designed activation functions, on the other hand, are typically well-reasoned and supported by theory or hypotheses, allowing them to generalize effectively across tasks and networks. However, the searched activation functions are difficult to explain in terms of why they exhibit certain characteristics, and they lack generalizability.

Questions

Weaknesses 2 and 3 raise concerns regarding the significance and necessity of the proposed problem and solution. If the authors are unable to address the issues of generalizability, I would be inclined to view their "improvement" as overfitting to specific tasks and networks.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

1 poor

Limitations

No potential negative societal impact.

Reviewer UeiT2023-08-14

Update after rebuttal

It is unfortunate that the authors' rebuttal did not address my concerns. 1. Firstly, the author's response did not effectively address the concern regarding overfitting. If the functions found during the search on a particular task or model cannot be generalized to other tasks or models, then it constitutes a form of overfitting. This so-called "customization" lacks practical significance and does not offer new insights for academic research. 2. Secondly, taking into account the opinions of Reviewer N57C and Reviewer x7yQ, I am more inclined to agree with N57C. The improvement brought about by this costly "customization" is extremely marginal. I will keep the rating as 3. Reject.

Authorsrebuttal2023-08-15

**Response to Reviewer UeiT** Thank you for taking the time to respond. We strongly disagree with your assessment and have responded to each of your points below. First, stating that the functions “cannot be generalized to other tasks or models” is a complete misrepresentation of the paper. Table 2 provides a direct contradiction to this statement: It shows that all nine of the activation functions discovered successfully generalized to a new task: ResNet-50 on ImageNet. Second, stating that customization is “overfitting” and “lacks practical significance and does not offer new insights” is patently false. Developing custom activation functions for better performance on specific tasks is something that human researchers regularly do, and this paper provides a way to automate this design process. Here is a concrete example: when modeling higher-order derivatives of a signal, periodic activation functions perform exceptionally well, while traditional activation functions like ReLU fail (https://arxiv.org/abs/2006.09661). Designing an activation function with the task in mind does not constitute overfitting! Similarly, one would not argue that CNNs have overfit to vision tasks or that RNNs have overfit to language modeling. Rather, these are models designed to exploit task-specific structure in the data. Our contribution is an automated method for designing activation functions that can also exploit task-specific structure to achieve better performance. Third, we strongly disagree that the performance improvement is “marginal.” Our approach provided a full percentage point increase in accuracy over ReLU on four different tasks (Tables 1 and 2). This performance improvement is on par with other work in the literature, and it is substantial given that so much effort has already gone into optimizing models for CIFAR-100 and ImageNet. Again, we appreciate your time in reviewing our paper, but many of the points you made contradict the facts in the paper. Thus, we hope you that will reconsider your point of view.

Reviewer N57C4/10 · confidence 3/52023-07-11

Summary

This paper introduces a set of benchmark datasets for activation function search, and an efficient search method based on the analysis of the benchmarks.

Strengths

1. The proposed benchmark datasets are beneficial for further research. 2. The method that searches activation functions through the function outputs of a limited number of samples seems effective and can significantly outperform the random search baseline.

Weaknesses

1. The paper is hard to follow. The main text refers to many details in Appendix, but it is still complex and hard to get to the method. I suggest the authors refine the structure and make the technical details of the proposed method more clearly. 2. Analysis is limited to show the efficiency of the method. The paper includes "Efficient" in the title, but I can only find the evaluation of efficiency in Appendix, and it should be compared with previous search methods to show how efficient it is. Besides, this method still needs to train multiple activation functions independently, which is also computationally expensive. 3. The improvements are marginal. The authors should compare their method with existing approaches in both benchmark search and new tasks search. Besides, in Table 1, comparing with the popular searched activation Swish, the improvements are marginal.

Questions

None

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

None

Reviewer x7yQ2023-08-16

I appreciate the Authors' response and clarifications. Incorporating these insights into the paper will definitely enhance its readability. Given the current state of the paper, I would keep my rating of 7. However, I believe demonstrating that the proposed method, including the choice of regression algorithm and embedding dimension 1) works on a model other than those used for the benchmarks, and especially that 2) the method can discover activation functions that outperform other baseline activations, even if by adding the baseline activation to the list of initial activations, would further demonstrate the strength of the method and improve the quality of the paper. Regarding the generalizability concern raised by reviewer UeiT, I respectfully hold a different perspective. Activation functions are part of the network architecture which can be tailored by human experts for a particular task, just like any hyperparameter which is optimized on a validation set, and therefore this shouldn't be considered as overfitting.

Reviewer N57C2023-08-20

Thanks for the response to my proposed questions. After reading them, part of my concerns are resolved. However, I do not see more competing results or explanations during the rebuttal phase. I am worried about the quite marginal improvement against the human-designed baselines, on which I agree with Reviewer UeiT and whether it is up to the standard of NeurIPS. Besides, apple-2-apple comparison in the empirical setting is important for the NAS community, which is also a weakness in this manuscript. With this regard, I tend to keep my original rating.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC