acceptodds
Under review as a conference paper at ICLR 2027

SEA-Net: A Small and Inherently Interpretable Neural Network for Time Series Classification

Abstract

Traditional time-series classification methods often operate as black boxes, providing a class prediction without indicating which time point or time interval within the series led to that prediction. To improve temporal interpretability, MILLET (Multiple Instance Learning with Temporal Explanation), which combines an InceptionTime encoder with Conjunctive Pooling, provides both classification and temporal explanations. However, MILLET contains approximately  K parameters, making it unsuitable for deployment on resource-constrained micro-controllers, and it achieves an overall time-series accuracy of and an explanation score of NDCG@. We therefore propose SEA-NET, a lightweight architecture using small multi-scale depthwise-separable convolutional TSC encoders and improved Multiple Instance Learning (MIL) pooling heads. Our bottleneck encoder with Top- pooling achieves accuracy and NDCG@ with only  K parameters. Another variant, using a gated encoder with Top- pooling, achieves accuracy and NDCG@ with  K parameters. These results demonstrate that lightweight architectures can maintain strong classification performance and temporal explanations while substantially reducing model size. The proposed variants address specific limitations of the pooling strategy used by , providing a practical direction for interpretable time-series classification on resource-constrained edge devices.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Analysis, Time Series Classification, Multi-Instance Learning, Convolutional Neural Network

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_iwhHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces SEA-Net, a lightweight and naturally interpretable model for classifying univariate time series. It extends the multiple-instance learning framework adopted by MILLET, where a time series is viewed as a bag and each time point is considered an instance. SEA-Net swaps out MILLET’s fairly large InceptionTime encoder for lightweight, multi-scale depthwise-separable convolutional encoders and investigates enhanced MIL pooling heads. The problem is meaningful, and the reported trade-off between model size, accuracy, and interpretability is encouraging. However, the current experiments do not sufficiently substantiate several key claims of the paper, particularly regarding suitability for microcontrollers, intrinsic interpretability, faithfulness, and overall performance on time-series classification.

Critical strengths
  1. The suggested architecture is technically well grounded, employing multi-scale depthwise-separable convolutions along with multiple-instance-learning pooling.

  2. The bottleneck and gated variants demonstrate a significant multi-objective trade-off between model size, accuracy, and the quality of explanations.

  3. Reducing MILLET from about 424K parameters to 41K–62K parameters represents a significant decrease.

Critical weaknesses
  1. The effects of the encoder and the Top-k pooling are not clearly disentangled.

  2. There are no compact baselines that are matched in terms of parameters and computation. The interpretation assessment may be strongly influenced by the use of synthetic temporal intervals.

  3. Microcontroller suitability is assessed without real-world deployment data, including latency, peak SRAM usage, energy consumption, or quantization outcomes.

Questions for authors

Q1: What are the peak SRAM, flash, MACs, latency, and energy?

Q2: Are the temporal explanations faithful under deletion and insertion tests?

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

This paper proposes SEA-Net, a light weight version of MILLET. MILLET is a time series classification algorithm that uses multiple-instance learning. The main motivation behind the paper is that the original MILLET cannot be used in memory limited environment (i.e., MCU) and the paper combines several existing memory efficient methods (e.g., squeezenet, depthwise conv) to make it run with smaller model weights while maintaining high performance. The experiment on TSC interpretability was conducted on WebTraffic, and baselines were mainly the variants of MILLET and the encoder structure.

Critical strengths

This paper proposes SEA-Net, a lightweight version of MILLET. MILLET is a time series classification algorithm that uses multiple-instance learning. The main motivation behind the paper is that the original MILLET cannot be used in a memory-limited environment (i.e., an MCU), and the paper combines several existing memory-efficient methods (e.g., SqueezeNet, depthwise conv) to make it run with smaller model weights while maintaining high performance. The experiment on TSC interpretability was conducted on WebTraffic, and baselines were mainly variants of MILLET and the encoder structure.

Critical weaknesses
  • The primary motivation of this paper is that MILLET, with 424K parameters, is unsuitable for resource-constrained MCUs. Accordingly, the authors propose SEA-Net, a light-weight version of MILLET, which is reportedly lighter (41K parameters) while achieving better performance. The reviewer believes that the current paper does not meet the standard for ICLR for five reasons.

(1) The application of the proposed model is limited to MILLET. While MILLET can be known to the time-series community, it is not a must-know algorithm in the time-series domain. As such, while the proposed model might be useful in specific sub-fields within the time series, the applicability of the proposed work is limited.

(2) The motivation is not supported by any on-device MCU experiment (Appendix E.2 is just an analysis for feasibility), and, to the best of the reviewer’s knowledge, it does not hold. Existing studies [1, 2] indicate that a model with 424K parameters can be deployed on an MCU. For example, [1] deployed a model with 6.26M parameters on a smartwatch without any quantization, and [2] deployed a model with 203K parameters (3.28 MB) that fits within 128 MB of MCU memory. To justify the need for SEA-Net, the authors should demonstrate that MILLET actually fails in an MCU (and in what particular MCU) environment, for example by reporting on-device measurements (e.g., memory footprint, inference latency on a target MCU). Moreover, even under a more constrained MCU setting, the memory issue could be readily mitigated by standard techniques such as quantization; since the paper assumes 32-bit precision, reducing it to 8-bit alone would shrink the model size by roughly a factor of four. The necessity of SEA-Net would therefore be convincingly established only if MILLET still fails to operate properly after such techniques are applied. Otherwise, the motivation should be revised accordingly.

(3) Regarding the motivation that interpretability is needed in MCU, it would benefit the paper to provide more concrete examples of where interpretability is needed for MCU. Currently, the reviewer believes that most of the interpretability analysis is conducted on the server-side, and is not familiar with why interpretability is needed at the MCU level. As such, the paper does not motivate this aspect enough.

(4) The proposed algorithm is a clever combination of existing algorithms (e.g., separable conv, squeeze and excitation modules; the paper acknowledges this). While the reviewer appreciates this honesty, the reviewer believes that this is not a fit for ICLR.

(5) The experiment is conducted on a limited number of datasets, and the main comparison is between MILLET and its variants. While the authors state that WebTraffic was the only dataset with known GT, there are plenty of other time-series data with known GT (e.g., MIT-ECG dataset). The authors should include more recent baselines and additional datasets for broader coverage.

[1] WatchHand: Enabling Continuous Hand Pose Tracking On Off-the-Shelf Smartwatches

[2] Smart ECU: Scalable On-Vehicle Deployment of Drivetrain Fault Classification Systems

Questions for authors

Please see above.

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

The proposed SEAnet is an optimization of MILLET by replacing its original InceptionTime encoder with a lightweight convolution-based encoder. They notably reduce the model params from 424K to 41K parameters.

Critical strengths
  • Motivation of deploying interpretable time-series classification models on edge/resource constrained device is relevant.
Critical weaknesses
  • The paper's contribution is extremely thing. It seems to be an extensive ablation and optmization of the already introduced work MILLET.

  • Overall the work is motivated by edge deployment but authors have not deployed on edge or profiled on resource contrained platforms for system-level metrics like latency, GFLOPs or MMACs to show its utility. Mere param count is not enough. Table 2 does report analytical flops but why not measured?

  • Overall evaliations ois very weak : No baselines from time-series interpretability field [R1, R2, R3, R4, R5]

  • If the goal is deploying on-device why not try pruning and other quantization methods?

[R1] Queen et al. Encoding Time-Series Explanations through Self-Supervised Model Behavior Consistency (TimeX). NeurIPS, 2023.

[R2] Liu et al. TimeX++: Learning Time-Series Explanations with Information Bottleneck. ICML, 2024.

[R3] Wen et al. Abstracted Shapes as Tokens--A Generalizable and Interpretable Model for Time-series Classification (VQShape). NeurIPS, 2024.

[R4] Wen et al. Shedding Light on Time Series Classification using Interpretability Gated Networks (InterpGN). ICLR, 2025.

[R5] Pandey et al. TimeSliver: Symbolic-Linear Decomposition for Explainable Time Series Classification. ICLR, 2026.

Questions for authors

See Weakness.

Reviewer_j6YXOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The major contributions seem to lie in replacing the backbone of of MILLET with one that is much more parameter efficient. This is done by using dilated and depthwise separable convolutions which have existed in the CV literature for quite a while as well as factorized pointwise mixing. A gating layer is explored. Additionally, the authors introduce a classwise and top-k conjuctive head. In general, this paper seems to present a series of architectural choices that seem grounded in CV literature to the 1-d problem of time-series classification. They all have merit but they do not seem to introduce any re-framing or new way of thinking about this problem that provides more insights. Prior work in this domain (MILLET, TimeMIL, InterpGN, TimeSliver) accepted at major conferences seems to have a flavor of re-framing how to address and model which then moves in the direction of ingenuity and I do not see any major contribution there to point to.

While these contributions generally reach into the domain of engineering, the authors do claim and have discussion of troubles with the AOPCR metric. This is highly relevant, but the framing of the paper does not seem to investigate this in a great deal of depth. Section 4.1 contains a paragraph which discusses the limitation and steps taken to address it, but it leaves myself as a reader desiring a more in-depth analysis, discussion. I was not able to find a more substantive discussion within the full 53 pages of the submission.

Critical strengths
  • The authors demonstrate clearly that with far fewer parameters, SEA-Net achieves strong results for accuracy and interpretability on WebTraffic as well as 85 UCR.
  • The authors point out flaws in the AOPCR metric that seem of high relevance to this field
  • The paper is clearly written and generally well-motivated
Critical weaknesses
  • The authors motivate the need for SEA-Net by deployment on edge devices; however, there are some weaknesses in motivation, results, and comparisons. As a reader I desire more motivation on why both interpretability and memory-efficiency are desired. Edge devices are clearly resource-constrained; however, it is unclear to me without more discussion or motivation that interpretability is important on-device as opposed to downstream. Furthermore, in the appendix Section E.2, the authors note that empirically SEA-Net's peak memory usage is actually higher than the baseline "179.0 MB against 112.3 MB for the baseline" and latency is only improved "1.6× even though the FLOPs fall 11.1×, because depthwise convolutions are limited by memory speed." It is commendable the authors did indeed obtain these results, and it is unfortunate they weaken the claims. For a paper motivated by the use case of running on edge devices, these empirical results weaken the pragmatic case for SEA-Net as it doesn't seem to offer as substantial memory or latency benefits that claims about FLOPs and numbers of parameters would imply. The authors note that GPU measurements are not representative of edge hardware; however, they do not follow this up with on-device experiments showing how performance would differ there.
  • For a paper to be motivated by this use case, I would desire results such as those in Section E.2 in the main manuscript and potentially experiments on hardware. However, in that case I would wonder if ICLR is indeed the right venue for a work focused on time-series classification on edge-devices. In a similar vein, I would desire other baselines such as an implementation of MILLET using fewer parameters, an attempt at using structured or unstructured pruning on a baseline such as MILLET or at least a discussion on why it would not be appropriate.
  • This work should also cite TimeMIL (ICML 2024), and TimeSliver (ICLR 2026) and either include as a baseline or discuss why they would not be appropriate baselines for comparison. InterpGN (ICLR 2025) is mentioned and I would also be interested in a comparison or discussion on why this would not be appropriate.
  • I generally had the impression reading this that the paper's writing was mostly done using a language model and it does not feel like an authentic human's writing. The writing is clear, but the register does feel ai-written, but it is not egregious.
Questions for authors
  1. Section E.2 shows SEA-Net's peak memory is higher than MILLET's (179.0 MB vs 112.3 MB). Can you discuss whether SEA-Net would be expected to reduce peak memory on edge devices, where SRAM is often the bottleneck?

  2. Are you able to benchmark SEA-Net on a representative edge device, along with baselines such as a smaller MILLET (fewer filters) and a pruned or quantized MILLET?

  3. Would you be able to discuss the merits of SEA-Net compared to TimeMIL, TimeSliver, and InterpGN, and either benchmark them or explain why they may not be appropriate?

  4. Can you discuss why interpretability is needed on-device as opposed to downstream?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.