acceptodds
Under review as a conference paper at ICLR 2027

Sensor Pre-Acquisition Neural Architecture Search

Abstract

Optimal neural architectures depend on input data acquisition, yet neural architecture search (NAS) typically fixes acquisition, excluding alternative acquisition–architecture pairs. Changing pre-acquisition parameters changes observations, preventing direct backpropagation through acquisition choices. Black-box joint search avoids this requirement but incurs substantial cost from training candidate pairs separately. We propose SPANAS, a NAS framework exploiting ordinal acquisition parameters to jointly search configurations and architectures without a differentiable acquisition model. A Rao–Blackwellized loss contrast between adjacent measured levels yields an unbiased within-interval selector gradient under fixed shared weights and generator parameters, without interpolating observations. A structured configuration representation conditions a hypernetwork over a weight-sharing supernetwork. Refinement and retraining verification produce an empirical accuracy–storage front. We evaluate SPANAS on FMCW-radar marshalling-signal recognition and a public wearable inertial benchmark. In both domains, SPANAS achieves higher mean hypervolume and mean best recognition performance than the evaluated black-box campaigns, raising hypervolume by 11% on radar and 7% on IMU over the strongest campaign. Its front also dominates every evaluated differentiable NAS result in both domains. On IMU, SPANAS achieves a higher macro-F1 than the strongest domain-specific architecture at the placement recommended by prior work with 5.7× less storage.

Primary area:
optimization
Keywords:
Joint Optimization, Neural Architecture Search (NAS), Multi-objective Optimization, Sensor System, On-device AI

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_1YMhOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

Summary: the paper performs a joint search over neural architectures and sensor settings, and shows that their search method can produce effective architecture-sensor setting combinations for movement recognition tasks.

Critical strengths
  • The method accounts for the geometry of the discrete ordinal space over which the search is conducted.

  • The specific problem of co-adapting sensor settings and architecture hasn’t been studied in the literature before to my knowledge, and the architecture-sensor location combination found does seem to outperform only tuning the architecture.

  • I didn’t spot any errors in the mathematical claims

Critical weaknesses
  1. The problem is not naturally motivated. It is true that nobody has tried to co-search over architecture and sensors, but this is perhaps because the problem doesn’t arise naturally, rather than because the problem is important and unsolved due to its difficulty.

  2. The Rao-blackwellized score estimation is, as far as I can tell, just finite differences. Finite differences is a very reasonable strategy for estimating derivatives of discrete points on the line, and the authors should be up front about the method rather than trying to obfuscate the approach behind unnecessary statistical machinery.

  3. The ‘co-design’ framing feels a bit inaccurate, since the method uses a two-phase process to first select a good set of sensors and then runs NAS with respect to those fixed sensors.

Questions for authors

Q1. How specific is the method to the data modalities considered in this paper? Could it be used for other types of feature selection?

Q2. Is W2 correct, or is there some advantage of the Rao-blackwellization that I missed which yields different properties from finite differences?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 1YMh (1/3)

Response to Weakness 1. Thank you for this comment. We agree that the submission did not explain the motivation sufficiently. To strengthen it, we revised the paper: the introduction now states which existing studies the motivation rests on, which measurements support it, and what the method contributes (the changed sentences carry the tag [1YMh W1]), and Appendix A has a new heat map of the 27 radar settings (Fig. 5).

Every sensing field already studies this problem, but one sensor at a time. Sensor pre-acquisition parameters must be set before any data are recorded, and they change the recognition performance. For this reason, published studies in each sensing field tune them, in applications from autonomous driving (LiDAR placement) to wearable health monitoring (IMU, PPG, EEG). Table 7 of the revised paper lists 28 such parameter families; some examples are below.

Sensing fieldParameter set before recordingPublished studies
FMCW radarframe period, bandwidth, antenna countKim & Kim (2024)
LiDARmounting positions, number of sensorsHu et al. (2022); Li et al. (2024)
Wearable IMUsensor placementAtallah et al. (2011); Tsukamoto et al. (2023)
Wearable IMUsampling rateKhan et al. (2016)
EEG, surface EMGelectrode subset and positionsStrypsteen & Bertrand (2021); Zhu et al. (2021)
Hyperspectral imagingspectral bandsLorenzo et al. (2020)
Cameraexposure, gain, processing parametersMosleh et al. (2020); Yan & Dansereau (2025)

Examples from Table 7 of the revised paper (Appendix B).

These studies show that the problem is important in each field. However, each of them handles one sensor with its own procedure, and none of them searches the architecture. The suitable setting, in turn, depends on the architecture (Enokibori et al., 2025, for IMU placement; our radar measurements are in (ii) below), so the setting should not be tuned without regard to the architecture.

In short, our core contribution is a search in which the sensor pre-acquisition parameters are optimised under the influence of the model architecture: no gradient passes through the sensor, yet the loss of the architecture drives the selection of these parameters. To our knowledge, SPANAS is the first gradient-based search that optimises sensor pre-acquisition parameters in this way, and the first method that does so across sensing domains with one procedure: the same method selects radar chirp parameters, IMU placements and speech acquisition parameters (Table 1). Three measurements show that ignoring this dependence is costly.

(i) The accuracy that can be reached changes a lot with the sensor setting: from 0.43 to 0.86 across the 27 radar settings (Fig. 5).

(ii) Which architecture is suitable differs from one sensor setting to another: different architectures reach their best accuracy at different settings. Table 6 lists ten architectures whose best setting is not the overall best setting, 160/255/30. Fixing the setting to 160/255/30 in advance costs these architectures up to 0.26 accuracy. If the two choices were independent, this loss would be zero; because they are not, the setting cannot be chosen without regard to the architecture.

(iii) SPANAS, which selects the setting with the loss of the architecture, reaches a comparable accuracy with 5.7× less storage than the strongest IMU model from the literature at its recommended placement: 0.857 ± 0.007 at 232 KB against 0.860 ± 0.036 at 1,325 KB (Table 2).

Searching the setting as well as the architecture does not mean searching the product of the two spaces. The selector moves along each ordered setting axis by comparing neighbouring settings, and the generator supplies an architecture suited to each setting, so the cost grows with the number of axes.

Where in the revised paper: §1 (new sentences, tagged [1YMh W1]: the existing studies of each sensing field, our measurements, and the missing method; contribution list); Appendix A (Fig. 5, new; Table 6); Appendix B (Table 7, now referenced from §1); Table 2 (§4.3).

AuthorsOpenReview ↗

Response to Reviewer 1YMh (2/3)

Response to Weakness 2 and Q2. Thank you for this careful reading. We agree with the reviewer: the selector update is a finite difference. For each setting axis, the update is the difference between the losses measured at the two neighbouring levels (Eq. 6). Rao–Blackwellization gives this finite difference no property that a finite difference does not have; it is only the route by which we derive it. The selector rounds each axis at random to one of two neighbouring levels. The usual single-sample (score-function) gradient of this random rounding is noisy, and averaging that gradient over the rounding in closed form leaves exactly the difference of the two neighbouring losses. The derivation tells us two things about the finite difference: it is an unbiased gradient of the expected loss, and its variance is no larger than that of the single-sample gradient (Proposition 1). The row "w/o finite difference" of Table 4 replaces the finite difference with the single-sample gradient:

Variant (Table 4)Radar HVIMU HV
Full (SPANAS), finite difference0.3250.263
w/o finite difference (single-sample gradient)0.3250.190

We agree that the submission should have said this up front. The paper now calls the update a finite difference throughout: in the abstract, §1, §3.2, the ablation of §4.4 (the row is now named "w/o finite difference") and the conclusion. Rao–Blackwellization is named only in the related work (§2) and in the derivation (Appendix D.1). What is new is the use of this finite difference: it gives the loss of the architecture a gradient with respect to sensor settings that are separate measurements, which, to our knowledge, no earlier NAS method obtains.

Where in the revised paper: Abstract; §1; §2; §3.2; §4.4 (Table 4); §5.

Response to Weakness 3. Thank you for this comment. We understand that the word 'co-design' can be read as optimising the two at exactly the same time and in one procedure, whereas the method has two phases. To avoid this reading, we removed 'co-design' and 'joint' as descriptions of SPANAS from the revised paper (abstract, introduction, contribution list, section title, Algorithm 1 and the tables).

The contribution of the paper is that the sensor pre-acquisition parameters and the architecture parameters are optimised in a coupled way, each with the current state of the other. In Phase A the two are updated in the same loop: for the current sensor setting a network proposes an architecture, the data acquired at that setting pass through this architecture, and the measured loss updates both the selector of the setting and the network that proposes the architecture (§3.2, Eqs. 4 to 6). Thus the loss of the architecture drives the selection of the sensor pre-acquisition parameters, and the architecture follows the selected setting. Phase B then refines the architecture at the selected setting. To our knowledge, this coupled optimisation is the first of its kind: a gradient that runs end to end from the architecture search to the setting of the sensor pre-acquisition parameters has not been proposed so far for sensors without a differentiable acquisition model, and existing methods cannot obtain it, because each setting is a separate measurement. SPANAS builds this gradient from measured losses.

This makes three things possible that existing methods do not offer. (i) The sensor setting is selected with a gradient although the sensor is not differentiable; differentiable NAS can only relax choices inside one recorded input. (ii) One shared network is trained once; black-box search trains every candidate pair of setting and architecture separately. (iii) The architecture is generated for the setting that is selected, so the two fit each other.

The ablation supports this. If the network that proposes the architecture is not told the setting, Phase A selects a different and less accurate sensor placement on IMU (Table 4, "w/o conditioning": IMU HV 0.263 → 0.162). On radar, with 27 settings, both variants select the same setting, so the HV is the same.

Where in the revised paper: Abstract; §1 (text and contribution list); §3.2 (new sentence); §3.4 (title, Algorithm 1); §4.3 (Table 2); §5; Fig. 1. The evidence is in Table 4 (§4.4).

AuthorsOpenReview ↗

Response to Reviewer 1YMh (3/3)

Response to Q1. Thank you for this question; it led us to add a third modality, speech: the 35 spoken words of the public Speech Commands v2 benchmark (Warden, 2018). The same recordings are re-acquired under 81 settings. Sampling rate, ADC bits and recording window are ordered settings, and the ADC encoding curve (linear, μ-law, A-law) is an unordered setting. The same selector handles both kinds: on an ordered setting it compares the two neighbouring levels, and on the unordered setting it compares all options. Results (Appendix I):

(1) SPANAS selects 16 kHz, 1 s, μ-law in all three seeds. Its selections rank 2nd, 2nd and 3rd of the 81 settings, within 0.004 accuracy of the best setting (Fig. 11). The rank comes from a separate reference run: 12 reference architectures were trained at each of the 81 settings, and the settings are ordered by the best test accuracy reached there.

(2) The HV of SPANAS is 0.402 ± 0.003, at the same level as the four black-box baselines (0.397 to 0.403; Table 17). The whole SPANAS pipeline takes less time on the same machine (one NVIDIA RTX 5090 GPU, every run alone): 20 minutes on average, against 31 to 39 minutes on average for a black-box baseline, counting its search and the verification of its front (Fig. 4c). The reason is that a black-box baseline trains 100 candidate models one after another, whereas SPANAS trains one shared network and then retrains only the few cells of its shortlist.

MethodVerified HV, seed 0Verified HV, seed 1Verified HV, seed 2Mean ± stdBest acc.Best acc. ≤ 401 KBFront points (seeds 0/1/2)
Random search.395.405.409.403 ± .007.916.7896/4/5
NSGA-II.385.410.401.399 ± .012.920.7935/5/5
qNEHVI.392.401.398.397 ± .004.904.8624/5/3
qNParEGO.401.400.409.403 ± .005.915.8495/3/4
SPANAS (six weights).403.404.399.402 ± .003.916.8632/4/3

Table 17 of the revised paper: speech, verified HV per seed and mean ± std, best accuracy, best accuracy among cells of at most 401 KB, and the number of points on each verified front.

Where in the revised paper: Abstract; §1; §4.1; §4.2 (Table 1, Fig. 4c); §5; Appendix I.

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

The authors propose a neural architecture search framework that jointly optimizes the neural network architecture and the sensor acquisition parameters, without requiring a differentiable acquisition model. They observe that acquisition parameters are largely ordinal, making standard Gumbel-Softmax relaxation less appropriate because it does not preserve the underlying ordering information. To address this limitation, the authors introduce a Rao-Blackwellized loss contrast between adjacent configuration levels. Experiments on radar and inertial wearable benchmarks show that the proposed approach improves hypervolume and best recognition performance over NAS baselines.

Critical strengths

[S1] The paper tackles a practically relevant problem and provides an end-to-end pipeline that jointly searches pre-acquisition parameters and neural architectures, offering a useful framework for practitioners.

[S2] The method is evaluated on two different sensing domains (FMCW radar and wearable IMU) and compared against both black-box multi-objective search methods and differentiable NAS baselines, as well as domain-specific architectures. The evaluation also includes extensive component ablations and additional results in the appendix.

Critical weaknesses

[W1] Argmax-based selection and lack of diversity in the front. Differentiable NAS methods typically learn a distribution over architectural choices (e.g., via Gumbel-softmax) and then derive a set of candidate architectures from it. SPANAS follows the same scheme: it discretizes the configuration, s∗=discretize(θ)\mathbf{s}^\ast = \mathrm{discretize}(\boldsymbol{\theta}) and builds the candidate set C\mathcal{C} from the top-kk cells of the arg⁡max⁡α\arg\max \boldsymbol{\alpha} trajectory. However, nothing in this procedure actively promotes diversity among the resulting solutions.

First, all architectures on the front of a single search share the same pre-acquisition configuration s∗\mathbf{s}^{\ast}. Second, the optimization of architectural logits through gradient descent does not encourage diversity. It optimizes the scalarized architecture loss of Eq. 8, which combines the validation loss with the expected storage cost weighted by a single fixed λ\lambda. The candidates are therefore close variants around one trade-off point rather than a spread along the accuracy–storage trade-off. In this case, a front would be obtained by running the search for several values of λ\lambda, which controls the accuracy-storage trade-off.

This also makes it questionable to present the results as a front. The method searches for a single trade-off point and the other points on the reported front are a by-product of the shortlist rather than the outcome of a multi-objective search. Presenting these solutions as a front and comparing them through hypervolume and GD against black-box multi-objective methods, which explicitly aim to cover the entire trade-off, seems a little bit off.

[W2] Unfair budget comparison in Sec. 4.2. Table 1 and Fig. 4 compare SPANAS with the black-box campaigns, but the two sides run under different budgets. The campaigns are limited to N=100N=100 candidates, each trained from scratch. SPANAS is instead limited by supernetwork search steps (Table 9) plus up to k=20k=20 retrained candidates. Fig. 4 puts both on the same "evaluated architectures" axis, but for SPANAS only the verified candidates count and the supernetwork training is left out. Table 12 is the only cost-matched comparison and it shows SPANAS needs more wall-clock time than qNParEGO and qNEHVI on IMU. A fair evaluation would compare the methods at matched compute (hypervolume against GPU-hours for example).

Questions for authors

[Q1] Related to [W1]: could the authors evaluate a variant that retains multiple high-probability configurations under pθp_{\boldsymbol{\theta}}, and compare the resulting front against the black-box baselines?

[Q2] A sigmoid is used to map θq\theta_q to rqr_q. Since its derivative mq(1−mq)m_q(1 - m_q) vanishes as mq→0m_q \to 0 or mq→1m_q \to 1, an axis whose rqr_q drifts toward an extreme rank may receive very small gradients (Eq. 6) and become hard to escape, even if the adjacent-level contrast favors another level. How sensitive is the selector to this and did the authors evaluate it?

[Q3] Random search performs quite well for both radar and IMU, outperforming the majority of the baselines in HV. Do the authors have an idea why?

[Q4] Table 1 reports the mean and standard deviation over three search seeds. Tables 2 and 3 instead report pooled or best-of-three results without any dispersion measure. Could the authors provide the variation for both the baselines and SPANAS in these tables as well?

[Q5] The authors compared wall-clock time against the black-box search algorithms (Table 12), but did not report it for comparable differentiable NAS algorithms. What is specifically the measured overhead of the contrast loss compared to the differentiable NAS methods in Table 3?

9 comments
AuthorsOpenReview ↗

Response to Reviewer fGgB (1/3)

Response to W1. Thank you for this question; the submission did not explain where the front comes from. We added the explanation (§3.4) and the cells that Phase B actually visits (Appendix C.3, Table 8). SPANAS optimises one objective that combines accuracy and storage, so it solves a scalarised multi-objective problem. Phase B does not go straight to one final cell: it fixes the six connections of the cell one per round, and until a connection is fixed it keeps switching between a 3×3 convolution, a 1×1 convolution, pooling and no operation. We keep the cells that Phase B visits in this way and retrain all of them; the most accurate cell at each storage level forms the front. This works like a grid search over the storage levels: in every search, the cells on the verified front lie at two or three different storage levels (284 to 1,343 KB; Table 8), not around one trade-off point. We observed that this gives SPANAS an advantage on the multi-objective measures, so we compared it with multi-objective optimisation methods (Table 1; highest HV and lowest IGD in both domains), and also with domain-specific models (Table 2) and single-objective NAS methods (Table 3). Running the search for several values of λ, as the reviewer suggests, changes little (IMU: HV 0.256 for the union of five searches against 0.247 for one; radar: 0.343 against 0.342; Appendix C.3).

Where in the revised paper: §3.4; Appendix C.3 (Table 8; the sentence on several values of λ). The comparisons are in Tables 1, 2 and 3 (§4.2, §4.3).

Response to W2. Thank you for pointing this out. We agree with the reviewer and have corrected the comparison. In the submission, the axis of Fig. 4 counted only our retrained candidates and left out the training of our shared network. Fig. 4 now plots HV against wall-clock time measured in the same way for every method: the method's own search plus the retraining of its own front. The table behind the figure (Appendix F.3, Table 15) replaces the search-time table of the submission (Table 12 there).

Even with time measured in this way, the results show the quality and the efficiency of SPANAS in the different domains. On radar, SPANAS takes 1.3–1.7× the time of the black-box baselines and reaches an HV that is 11% higher. On IMU, SPANAS is faster than these baselines (about 3.7 h against 4.1–9.9 h) and its HV is higher. On speech, the domain added during the discussion, SPANAS reaches the HV of the baselines in 20 minutes against 31–39 minutes (Fig. 4c).

Where in the revised paper: §1; §4.1; §4.2 (Fig. 4); Appendix F.3 (Table 15).

Response to Q1. Thank you for the suggestion; we tried it. We ran Phases B and C also at the setting that the selector ranks second and merged the two fronts. In all three radar seeds this adds 0.000 HV: the second setting adds no point to the front. We agree, however, that the variant you suggest, which keeps several settings, can make the search safer in other environments, for example when the best settings are close to one another. The number of settings that Phase A passes on is therefore a parameter of the method. Although we used one in our experiments, Appendix C.3 now introduces the variant with several settings and how to decide when to use it.

Where in the revised paper: Appendix C.3.

AuthorsOpenReview ↗

Response to Reviewer fGgB (2/3)

Response to Q2. Thank you for raising this; we analysed it and checked it on the logged searches (Appendix F.4, new Fig. 10). The derivative mq(1−mq)m_q(1-m_q) does shrink near the ends, but the update of the selector does not vanish, for two reasons. First, the finite difference measures both neighbouring levels at every update, also at an end, so the direction of the update is always available. Second, the selector parameter is updated with Adam, whose step is the running mean of the gradient divided by the square root of the running mean of its square. The derivative changes slowly between updates and multiplies both, so it cancels: the step depends on how consistently the comparison of the two neighbouring levels favours one side, not on the size of the derivative. Leaving an end therefore takes a number of updates proportional to ∣θq∣|\theta_q|; with plain gradient descent it would be proportional to 1/(mq(1−mq))1/(m_q(1-m_q)). The cancellation holds while the gradient is above the constant ϵ=10−8\epsilon=10^{-8} of Adam; in the radar searches of Table 1 the smallest derivative is 6×10−46\times10^{-4}.

The logged searches confirm this. (a) In radar seed 2, the ADC-samples coordinate rose to 0.97, where the derivative is 13% of its maximum, and was back at the middle 300 updates later. (b) Between the middle and the ends the derivative falls by a factor of 49 (IMU) and 65 (radar), while the net movement of θq\theta_q per update falls by a factor of 1.6 and 3. A numerical check of the update rule covers deeper saturation (loss difference with a mean of 0.15 standard deviations, 20 runs): from mq=0.982m_q=0.982, 0.9997 and 0.99995, Adam returns to the middle within about 640, 1,700 and 1,700 updates, whereas plain gradient descent with the same step at the middle needs about 2,000, 84,000 and 617,000. A coordinate stays at an end only when the comparison keeps favouring the end level (radar: chirps → 255, the best level; Fig. 8). If a guarantee is wanted, θq\theta_q can be clipped, which bounds the derivative from below; our runs did not need it.

Where in the revised paper: Appendix F.4 (Fig. 8; new paragraph "Saturation of the sigmoid" and Fig. 10).

Response to Q3. Thank you for this question. Reviewer u1Fp asked a similar question (Weakness 2), and we give the same explanation here. We spent more budget on the two methods in question: ten search seeds each for qNEHVI, random search, and qNEHVI with a larger initial design (new Table 14, Appendix F.2). Over ten seeds on radar, HV is 0.262 ± 0.080 for random search, 0.241 ± 0.105 for qNEHVI, and 0.252 ± 0.067 for qNEHVI with the larger initial design.

Random search does well because, in these runs, the guided search of qNEHVI gains little from its guidance when it chooses the sensor settings. Table 14 lists the three sensor settings that each run evaluated most often, with the rank of each setting among the 27 settings (the rank orders the settings by the best accuracy that an architecture search reaches at that setting alone, Table 5). Random search spreads its 100 evaluations evenly: its three most-evaluated settings take 18 to 23 evaluations. qNEHVI concentrates on a few settings: its three most-evaluated settings take 27 to 40 evaluations, and in eight of the ten seeds the most-evaluated setting is ranked 11th or lower. Overall, 67% of the evaluations of qNEHVI go to settings ranked 10th or lower, the same as random search (68%). So, in these runs, qNEHVI does not steer its search towards good settings noticeably better than random sampling does, and a run that concentrates on a poor setting ends below random search.

This agrees with the literature. Our search space combines discrete choices of different kinds, the sensor setting and the six operations of the cell, and categorical and mixed search spaces are reported to remain challenging for Bayesian optimisation (Wan et al., 2021; Daulton et al., 2022); methods designed for such spaces report clear gains over the standard approach (Ru et al., 2020; Deshwal et al., 2021). Random search is also known to be a strong baseline in architecture search (Li & Talwalkar, 2019). The revised paper now gives this explanation in §4.2 and Appendix F.2.

Where in the revised paper: Appendix F.2 (Table 14); Appendix F.4 (Fig. 9).

AuthorsOpenReview ↗

Response to Reviewer fGgB (3/3)

Response to Q4. Thank you for this request; we now report it. Tables 2 and 3 of the revised paper give, for every model (the baselines and SPANAS), the mean and the standard deviation over three seeds: every model is retrained three times.

With three retrainings, iSPL-Inception (the IMU model from the literature in Table 2) reaches 0.860 ± 0.036 at 1,325 KB and SPANAS reaches 0.857 ± 0.007 at 232 KB, so we now write "comparable accuracy at 5.7× less storage". Against the architecture-only control of Table 2, SPANAS is ahead from 96 KB upward, and the two are within one standard deviation at the smallest sizes (0.796 at 44 KB and 0.785 at 13 KB); the text now says this. Every differentiable-NAS row of Table 3 stays below a SPANAS model of equal or smaller storage.

Where in the revised paper: Abstract; §4.3 (Tables 2 and 3 and the text next to them); Appendix E (how the rows are built).

Response to Q5. Thank you for asking; we now report it in Table 3, which has a new column with the time that each method needs to select the sensor setting. The differentiable NAS methods that are extended to the acquisition axes run in the same pipeline as SPANAS, with only the selector of Phase A replaced, so the overhead of our comparison of neighbouring settings is the difference in the time of Phase A. On radar, where every run was repeated alone on one GPU, Phase A takes 43 minutes with our selector, against 18 with Gumbel-ST, 36 with ProxylessNAS, 52 with DrNAS and 53 with the DARTS mixture. Phases B and C, which follow, are the same procedure for all of these rows. For SPANAS, the whole pipeline takes 325 minutes on radar (Table 15: 203 for the search, that is, Phases A and B, and 122 for the retraining of Phase C). The comparison of neighbouring settings therefore adds 25 minutes over Gumbel-ST, the fastest selector, which is 8% of the whole pipeline. The BB + NAS rows first need a black-box campaign, 166 minutes on average, before their NAS starts; the NAS at the selected setting then took 38 to 39 minutes for each of the four methods in the runs of Table 3.

Our selector is slower than Gumbel-ST because it evaluates the current setting and its neighbouring settings at every selector step (28 forward passes on radar, 72 on IMU, no backward pass), whereas Gumbel-ST evaluates one sampled setting. This extra work is what lets the selector converge: with Gumbel-ST the search ends on the settings ranked 3rd, 13th and 11th and reaches 0.640 accuracy, against 0.909 for SPANAS (Table 3, Fig. 8).

Where in the revised paper: §4.3 (Table 3, time column); Appendix F.4.

Response to Presentation. Thank you for the feedback on clarity. We read the method section again as a first reader would and rewrote the places that were hard to follow. (1) §3 now opens with what each of the three phases does. (2) §3.2 and §3.3 each start with a description in plain words before the equations: how the selector moves between two neighbouring levels, and how a setting is described to the network that proposes the architecture. (3) The update of the selector is now described directly as a finite difference between two adjacent measured levels. (4) §3.4 has a new paragraph that explains where the accuracy–storage front comes from, and Appendix C.3 lists the cells that the refinement keeps (Table 8). (5) We unified the terms; for example, "pre-acquisition parameter" is now used throughout the paper.

Where in the revised paper: §3 (opening); §3.2 and §3.3 (opening sentences); §3.4; Appendix C.3 (Table 8).

Reviewer_u1FpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose a NAS framework exploiting ordinal acquisition parameters to jointly search configurations and architectures without a differentiable acquisition model. It uses an ordinal selector with Rao–Blackwellized loss differences to update sensor configuration probabilities, and a hypernetwork to generate architecture distributions conditioned on different configurations. The method includes a joint search stage, architecture refinement, and independent retraining. Experiments on radar gesture recognition and IMU-based activity recognition show improvements over several black-box optimization and NAS baselines in terms of accuracy and model storage trade-offs.

Critical strengths

The problem is practically relevant and the method is clearly presented.

Critical weaknesses
  1. limited novelty. The method mainly combines existing NAS and gradient estimation techniques. The main new part seems to be using the loss difference between adjacent sensor configurations to update selection probabilities, which is rather incremental.

  2. qNEHVI performs much worse than Random Search on radar, with a very large variance. The paper does not fully explain why this happens.

  3. Missing comparisons with closely related Data-Aware NAS methods, e.g., Njor et al. (2025) and Shi et al. (2025). All sensor configurations are collected beforehand, so the actual optimization problem is not fundamentally different from existing joint data-configuration and architecture search.

  4. The need for Phase B is not clear. Phase A already learns a configuration-conditioned architecture distribution, but Phase B discards it and starts NAS again from scratch. The current ablations do not really answer this, since Phase C involves independent retraining. A direct comparison of A+C and A+B+C would be more informative.

Questions for authors

no further question.

5 comments
AuthorsOpenReview ↗

Response to Reviewer u1Fp (1/3)

Response to Weakness 1. Thank you for this candid assessment. We agree with the reviewer that the selector is updated by the loss difference between adjacent levels. However, our core contribution is a search in which the sensor pre-acquisition parameters are optimised under the influence of the model architecture: no gradient passes through the sensor, yet the loss of the architecture drives the selection of these parameters. The adjacent-level comparison is one part of this method. To our knowledge, SPANAS is the first gradient-based search that optimises sensor pre-acquisition parameters in this way.

What we observed. As a reference, we ran an architecture search with qNEHVI (100 evaluations) at each of the 27 radar sensor settings alone, and we ordered the 27 settings by the best accuracy found there (Table 5). In these runs, which architecture is suitable differs from one sensor setting to another (Table 6). Existing differentiable selectors treat the levels of a sensor pre-acquisition parameter as unordered categories. When we extended these selectors to the sensor pre-acquisition parameters and applied them in the same pipeline, they did not converge: the selection kept changing until the end of the search, and in its three seeds the categorical selector ended on the settings ranked 3rd, 13th and 11th among these 27 settings (Fig. 8, Table 3). This corroborates that no gradient passes through the sensor, because each change of a sensor pre-acquisition parameter changes the data themselves; to our knowledge, no existing differentiable search handles this case.

What we propose. From these observations we propose a method in which the loss of the model architecture drives the optimisation of the sensor pre-acquisition parameters. The selector chooses a sensor setting, and the generator, a separate network, proposes an architecture for that setting. The data acquired at that setting pass through this architecture. The resulting loss updates the selector, which compares it between two adjacent levels of each sensor pre-acquisition parameter, and it also updates the generator. With this method, the selector ends on the setting ranked 1st in all three seeds (Fig. 8).

Where in the revised paper: §1 contribution list; Appendix F.4 (Fig. 8, Fig. 9). The evidence is also in Tables 5 and 6 (Appendix A) and Table 3 (§4.3).

Response to Weakness 2. Thank you for pointing this out; the submission did not explain it. We spent more budget on these two methods: ten search seeds each for qNEHVI, random search, and qNEHVI with a larger initial design (new Table 14, Appendix F.2). The spread remains. Over ten seeds, HV is 0.241 ± 0.105 for qNEHVI, 0.262 ± 0.080 for random search and 0.252 ± 0.067 for qNEHVI with the larger initial design. In the three searches of Table 1, SPANAS selects the rank-1 setting (Fig. 8).

Table 14 shows the reason. It lists the three sensor settings each run evaluated most often, with the rank of each setting among the 27 settings (the rank orders the settings by the best accuracy an architecture search reaches at that setting alone, Table 5). qNEHVI puts much of its search into low-ranked settings: its three most-evaluated settings take 27 to 40 of its 100 evaluations (random search: 18 to 23), and in eight of the ten seeds the most-evaluated setting is ranked 11th or lower. Overall, 67% of the evaluations of qNEHVI go to settings ranked 10th or lower, the same as random search (68%). So qNEHVI does not steer its search towards good settings, and a run that concentrates on poor settings ends below random search. This agrees with the literature. Our search space combines discrete choices of different kinds, the sensor setting and the six operations of the cell, and categorical and mixed search spaces are reported to remain challenging for Bayesian optimisation (Wan et al., 2021; Daulton et al., 2022); methods designed for such spaces report clear gains over the standard approach (Ru et al., 2020; Deshwal et al., 2021).

The revised paper now gives this explanation in §4.2 and Appendix F.2.

Where in the revised paper: §4.2 (new sentence); Appendix F.2 (Table 14); §2 citations.

AuthorsOpenReview ↗

Response to Reviewer u1Fp (2/3)

Response to Weakness 3. Thank you for these references. We added both methods as baselines in both domains (Appendix G.1).

What the two methods search. Both methods optimise how data that are already collected are processed. Shi et al. (2025) optimise the parameters of the audio features for keyword spotting: the window size, the hop length and the number of mel filters of MFCCs. Njor et al. (2025) optimise the image resolution and the colour encoding for person detection. In both cases every candidate is computed from the same collected sample, the same audio clip or the same image.

Comparison. We ran both methods with three seeds and scored them in the same way as SPANAS. For Shi et al., the setting that their selector chooses receives the same refinement and retraining (Phases B and C) as in SPANAS. For Njor et al., the front of their 100-evaluation search is retrained with the same protocol (Phase C).

DomainMethodHV per seed 0HV per seed 1HV per seed 2Mean ± std over seeds HV↑Mean ± std over seeds GD↓Mean ± std over seeds IGD↓
RadarSPANAS0.3480.3250.3530.342 ± 0.0150.000 ± 0.0000.019 ± 0.032
RadarShi et al.0.0680.0760.1590.101 ± 0.0510.144 ± 0.1280.231 ± 0.031
RadarNjor et al.0.1020.1580.0000.087 ± 0.0800.243 ± 0.0810.247 ± 0.096
IMUSPANAS0.2600.2610.2210.247 ± 0.0230.002 ± 0.0040.005 ± 0.008
IMUShi et al.0.1950.2110.1720.193 ± 0.0190.093 ± 0.0750.100 ± 0.060
IMUNjor et al.0.1560.1130.1200.130 ± 0.0230.069 ± 0.0590.112 ± 0.020

Why they are lower on our problem. Our problem is not how to process data that are already collected. We optimise the parameters that affect the data before the data are collected (pre-acquisition), which is fundamentally different and harder, because the data of one setting cannot be computed from the data of another setting. For example, when the chirp slope of the radar or the position of an IMU sensor changes, the data themselves become different (Appendix B, Fig. 6). The existing methods do not address this kind of choice, so they do not work properly on our problem: both are below SPANAS in both domains (table above), and on radar the sensor settings that they select rank lower than the setting that SPANAS selects (Appendix F.4, Fig. 9). This is also why our problem differs from the existing joint search of data configuration and architecture that the reviewer mentions. To our knowledge, no existing NAS method, including these two, optimises parameters that act before the data are collected with a gradient from the loss of the architecture.

Where in the revised paper: §1; §2; §5; Appendix G.1 (Table 16); Appendix F.4 (Fig. 9). The change of the data with a pre-acquisition parameter is shown in Appendix B (Fig. 6).

AuthorsOpenReview ↗

Response to Reviewer u1Fp (3/3)

Response to Weakness 4. Thank you for suggesting this comparison; we ran it in both domains. Phase A can hand two things to Phase B: the weights of the shared network, and the architecture logits (the scores that say which operation each connection of the cell prefers). The Phase B of the paper restarts both. We compared three alternatives. A+C skips Phase B and retrains the architectures that Phase A visited most often. B′ starts Phase B from the weights and the logits of Phase A. B″ starts Phase B from the weights of Phase A only.

VariantRadar HV seeds 0 / 1 / 2Radar HV mean ± stdIMU HV seeds 0 / 1 / 2IMU HV mean ± std
Full pipeline0.348 / 0.325 / 0.3530.342 ± 0.0150.260 / 0.261 / 0.2210.247 ± 0.023
A+C (no Phase B)0.272 / 0.307 / 0.2410.273 ± 0.0330.267 / 0.190 / 0.2060.221 ± 0.041
B′ (weights and logits kept)0.281 / 0.236 / 0.2360.251 ± 0.0260.217 / 0.161 / 0.1700.183 ± 0.030
B″ (weights kept)0.281 / 0.311 / 0.2810.291 ± 0.0170.274 / 0.253 / 0.2140.247 ± 0.030

A+C is below the full pipeline in all three radar seeds and in two of the three IMU seeds. B′ returns a single architecture in all six runs and is below the full pipeline in all six. B″ reaches the level of the full pipeline on IMU but stays below it on radar.

We read these results as follows. A+C is lower because Phase A is trained to choose a sensor setting: its architectures are proposed while the setting is still changing, so they are not tuned to the setting that is finally selected. B′ returns a single architecture because the logits of Phase A have already settled on one cell: Phase B then has nothing left to explore, and one model cannot cover several sizes. B″ shows that the weights matter less than the logits. The weights of Phase A were trained across all settings, not only at the selected one. When the settings differ strongly from one another (radar accuracy ranges from 0.43 to 0.86 across settings), weights trained at the selected setting alone rank the architectures better.

So Phase B is needed for the multi-objective search: it is the stage that returns models at several storage levels, which form the front. Restarting both the weights and the logits, as the paper does, is the safe choice.

Where in the revised paper: §3.4; Appendix C.4 (Table 9).

Comments on the paper

3 comments
AuthorsOpenReview ↗

General response

We thank all reviewers for their careful reading. During the discussion period we ran every experiment the reviews asked for. All new runs use the same training and scoring as Table 1. The revised PDF marks every change in red, with the review point it answers, and each answer below names where the new material is in the paper.

What is new in SPANAS. To our knowledge, SPANAS is the first NAS method in which the loss of the architecture gives a gradient to sensor pre-acquisition parameters without a differentiable acquisition model. Existing differentiable NAS methods cannot obtain this gradient, because the data of two sensor settings are separate measurements, and black-box methods need one full training for every candidate.

One change of wording applies to the whole paper. Following Reviewer 1YMh (Weakness 2), the paper now calls the update of the selector a finite difference between the losses at two adjacent measured levels, and no longer a Rao–Blackwellized estimator. Rao–Blackwellization is named only in the related work and in the derivation in the appendix.

New experiments: (1) two data-aware NAS baselines (Shi et al., Njor et al.) in both domains; (2) ablations of Phase B; (3) ten search seeds each for qNEHVI, random search, and qNEHVI with a larger initial design; (4) trajectories of the selector during the search; (5) time measured in the same way for every method; (6) a third modality, speech, where an unordered (categorical) setting is also handled by the same method.

Terms. HV: hypervolume of the accuracy–storage front after every model on it has been retrained (higher is better). Phase A: the search that selects the sensor setting. Phase B: the search that refines the architecture at the selected setting. Phase C: the retraining of the candidate models. Sensor pre-acquisition parameter: a parameter of the sensor that is fixed before recording, for example the number of radar chirps or the position of a wearable sensor. Sensor setting: one combination of the levels of these parameters. Selector: the part of SPANAS that chooses the sensor setting. Generator: the network that proposes an architecture for each sensor setting. Campaign: one run of a black-box search baseline (qNParEGO, qNEHVI, NSGA-II or random search) with 100 evaluations.

AuthorsOpenReview ↗

Second revision uploaded: summary of changes

We thank all reviewers for the discussion. We uploaded a second revision; its changes are marked in blue, and the changes of the first revision stay in red. The main text remains within the page limit. The additions of this round are:

  • Storage weight (Appendix C.3; Tables 9–11, Fig. 7): the exact weights, the cells selected under each weight, and the corresponding radar check.
  • Variation over independent searches (Appendix G; Tables 25 and 26): three searches for every method of Table 3, and the participant splits (Appendix E).
  • Cost (Fig. 4; Appendix F.3, Tables 18–22): comparison by training compute, the devices used, and a breakdown of the radar time of SPANAS.
  • Selector map (§3.2; Appendix F.4, Tables 23 and 24): comparison of the sigmoid with a clipped linear map.

Details are in our replies to the individual reviewers.

AuthorsOpenReview ↗

Third revision uploaded: summary of changes

We thank the reviewers for helping us clarify the practical motivation and contribution of SPANAS.

We have uploaded a further revision clarifying the engineering need for acquisition selection and its relationship to architecture search. The abstract, introduction and related work now explain the existing sensing applications and recent platform developments that motivate our approach, and clarify its contribution: combining adjacent-level loss comparisons, configuration-conditioned architecture search and shared training in a procedure evaluated across three sensing domains.

Changes in this round are marked in purple; earlier revisions remain in red and blue. The main text remains within the page limit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.