acceptodds
Withdrawn from ICLR 2027

Learned Intensity Sensors Trade Separability Across Orientations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A learned sensor is frozen in hardware, so whatever it fails to record is lost to every decoder. We ask what such a sensor trades away, for intensity sensors, which record squared magnitudes of a coherent field under additive Gaussian noise. How well one separates two draws from a signal distribution, measured by the average Kullback-Leibler divergence between their measurement laws, is an inner product of two fourth moments, one of the sensor and one of the data, and equals the variance the data induce in its noiseless measurements. Averaged over rotations of the distribution, this separability is the same for every sensor with unit-norm measurements: learning redistributes it and cannot create it, although on one distribution a sensor may hold far more than isotropy. The excess is an exact overlap with the sensor's fourth-order defect, which the distance to a complex projective 2-design bounds only uniformly over a distribution class, not on the task at hand. Under a controlled rotation, a learned sensor that beats a quadratic-mask design by 2.6 dB in distribution loses to it by 0.8 dB at with retrained decoders, and a decoder-free overlap ranks this loss across configurations. In a matched family where only the fourth-order direction varies, the effect is graded with the alignment but small, and its sign follows the local Cram\'er-Rao bound rather than the separation. Two boundaries are as informative: the ordering reverses at low noise, and in coded diffraction imaging of phase objects no learned mask set loses more than  dB of separability.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
computational imaging, end-to-end optical design, learned sensors, phase retrieval, coded diffraction imaging, Fisher information, Cramer-Rao bound, complex projective 2-designs, distribution shift

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_kxRWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors claim that a learned measurement operator that maximizes the measurement information for a specific training distribution may not be optimal when the training distribution is mismatched with the true distribution. To analyze this effect, the authors introduce the separation, which quantifies how much variance a measurement operator can capture under unitary rotations of the training distribution. Experiments show that when the true distribution is a rotated version of the original distribution, having a fourth order statistic in the regularizer provides a more robust measurement operator, but under more general distribution shifts the separation-optimal sensor does not show many improvements.

Critical strengths

None

Critical weaknesses

The paper is poorly written and difficult to understand. Several indications, such as unnatural phrasing and overly complex mathematical explanations, give away the fact that it is fully AI generated, with no efforts to improve readability afterwards. Many esoteric concepts such as "Haar twirl" are mentioned briefly, not explained as to why they are relevant, and never brought up again. Other crucial terms like "isotropy" of different orders are not defined until much later and their definition has to be inferred.

The theoretical analysis also focuses on when the true distribution is a rotated version of the training distribution; yet this is essentially never the case in practice. The experiments also demonstrate that when the true distribution has some other distribution shift, the separation does not really help, making the result not so useful in practice.

The fourth order statistics, which is the main contribution as proposed by the authors, also do not provide extra diagnostic information over the second order statistics.

Questions for authors

What is the benefit of using the fourth order statistics rather than the simpler yet equally informative second order statistics, if it doesn't help in practice?

Reviewer_cHoeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper establishes an orientation-averaged conservation law for the pairwise KL separability of intensity measurements and investigates how learned sensors redistribute this separability across data orientations, including its relationship to reconstruction performance.

Critical strengths

The paper provides a clear theoretical distinction between the magnitude of a sensor’s fourth-order design defect and its alignment with a particular data distribution. The target-trained decoder protocol and matched-frame experiments help disentangle different sources of specialization.

Critical weaknesses
  1. Limited evidence of practical relevance. The pronounced trade-off appears mainly in synthetic experiments; in coded diffraction imaging, learned masks lose at most 0.11 dB of separability under the tested rotations. It remains unclear how significant this phenomenon is in practical imaging systems.

  2. Unclear value for predicting reconstruction quality. The proposed statistic performs no better than simpler second-order diagnostics in the main experiment and can give the opposite ordering to reconstruction performance in the controlled experiment. Its usefulness for choosing between sensors is therefore uncertain.

  3. Limited actionable design guidance. Reducing the design defect can worsen the local estimation bound, and the proposed regularizer does not demonstrate a clear practical advantage over simpler alternatives. The paper would benefit from showing how its analysis leads to better sensor designs or more robust reconstruction.

Questions for authors

see weakness

Reviewer_tL7hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates what learned intensity sensors trade away when the data distribution rotates. The authors define separation as the average pairwise Kullback–Leibler divergence between measurement laws, express it as an inner product of fourth-order moments, and prove that its Haar average over unitary rotations is sensor-independent. Learning therefore redistributes separability across orientations rather than creating it. Controlled rotation experiments demonstrate in-distribution gains and off-distribution losses, and examine the effect of a design-defect regularizer. Experiments in coded diffraction imaging further illustrate the limits of separation as a predictor of reconstruction performance.

Critical strengths

This paper focuses on the specialization of learned sensors under distribution shifts, distinguishing between the inherent limitations of the sensors themselves and the maladaptation of the decoders; it holds potential significance. The theoretical framework is clear; it characterizes separability and its rotational-average conservation relation through fourth-order moments, and provides substantial numerical verification. The control experiments were conducted with care, and the authors proactively reported limitations such as inconsistencies between the indicators and reconstruction performance, and sorting reversal under low noise conditions.

Critical weaknesses

While the theoretical validation is quite thorough, the practical utility has yet to be demonstrated convincingly. Given that the paper is framed within the context of learning-based sensing and computational imaging, it requires a concrete application scenario to show that the theory can shed light on real-world problems or effectively guide design. This is my primary concern regarding the paper. Although the coded diffraction imaging experiments provide simulations for practical applications, they fail to replicate the distinct degradation in separability observed in the primary synthesis experiments, nor do they establish a reliable link between theoretical metrics and reconstruction quality or out-of-distribution robustness. In the main experiment, simple second-order statistics exhibit similar predictive power; furthermore, once the second-order structure is held constant, the relationship between "separability" and the quality of reconstruction appears to reverse. It seems that the incremental value of this core metric remains unclear. Without a side-by-side comparison of reconstructed images, it is difficult to determine what kind of actual imaging degradation corresponds to a change in theoretical metrics.

Questions for authors
  1. Which specific practical imaging problem can the theory presented in this paper help to solve? Could you provide a scenario where this theory is used for diagnosis or design and yields clear benefits? This is the question I am most keen to see the authors address.
  2. The coded diffraction imaging experiments did not replicate the significant degradation observed in the primary synthesis experiments. How do the authors interpret this discrepancy, and what are its implications for the claims made regarding the paper’s potential applications?

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

Related papers

Discussion (0)

Sign in to comment.