acceptodds
Withdrawn from ICLR 2027

Propagation Is Not Decoration: When Physical Diffraction Carries Generative Computation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

14 views

Abstract

We introduce the conditioned Diffraction Field Model (cDFM), a generative architecture built from Wave-Aperture-Diffraction (WAD) blocks. Each block conditions a complex field with a learned, class-conditioned aperture and propagates it with a fixed Angular Spectrum Method (ASM) operator. One random field is carried through four resolutions and read out directly as intensity, with no learned decoder. A learned aperture can form images by itself and bypass propagation, so the presence of propagation in such a model does not show that it does any work. We test this by retraining eight controls from scratch with matched parameter counts and update budgets. In an earlier image-fitting model, restricting the aperture to 16x16 spatial resolution lets ASM remove 30.7% of the excess error of a retrained bypass, compared with 2.1% at 64x64. On MNIST, the full generator reaches 83.1% class accuracy from fresh sources across three seeds. Replacing ASM with the identity, E(x)=x, drops this to 20.3%, and rebuilding the source at each scale drops it to 16.4%. An identity control whose aperture can amplify light reaches 78.5%, so the identity drop depends on the aperture being passive, as a physical aperture is. A nonphysical operator matched to ASM's local reach reaches 77.0%, and a global permuted spectral operator 55.7%. A wider version (994k parameters) trained for 30 epochs reaches 95.9% on one seed. Retrained at that budget, the passive identity control reaches 90.3% and the permuted spectral control 94.1%, but both fit held-out digits worse. In a separate strict-locality task, identity evolution leaves distant prediction at chance (25%), ASM at propagation distance 20 reaches  55%, and a global random unitary and a permuted spectral operator reach  86% and  88%. Propagation matters most in this generator when the aperture is passive and the learned path is small or briefly trained, and less as the learned path grows. In the locality task, diffraction is one of several global operators that can carry distant information. On CIFAR-10 the same design produced mostly blobs with a coarse lattice texture, so the generative results are limited to MNIST.

Primary area:
generative models
Keywords:
generative modeling, physics-informed models, optical computing, diffraction, angular spectrum method, ablations

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_QghSOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether propagation plays an important role in diffractive neural networks for image generation. The authors constructed a cDFM and conducted ablation studies by removing or modifying the physical propagation process. The experiment shows that propagation is not just an optional feature. Rather, it provides essential computational capabilities under certain constraints, such as if an aperture is passive or there is limited training time. In conclusion, this study argues that propagation is a fundamental mechanism that enables certain computational capabilities.

Critical strengths
  • The paper provides insights that propagation plays a meaningful computational role in diffractive generative models, and how this depends on model capacity and physical constraints.
  • The paper identifies the functional necessity of different components through ablation studies, which help to improve understanding of their contributions to the model's performance.
Critical weaknesses
  • W1. (Motivation) The motivation for removing propagation is somewhat unclear. Since propagation is an inherent physical process in optical systems, it is not clear what additional insight can be gained by removing ASM. As it is trivially expected that replacing ASM with an identity operation would reduce the model's representational capacity, the observed performance degradation alone may not be sufficient to demonstrate the importance of diffraction.
  • W2. (Analysis) While ASM may provide a useful inductive bias that improves learning efficiency, the recovery of the Identity model with increased network capacity and training budget is not sufficiently explained. In particular, it is unclear how the Identity model recover performance without propagation or light amplification.
  • W3. (Writing Issue)
    • The abstract and introduction mainly list experimental results rather than explaining their implications and key insights. As a result, it is difficult to understand the significance of the findings.
    • The related work section does not provide sufficient context to understand the position of this paper relative to existing studies.
    • Several mathematical notations(R_s,σ, ℜ... Eq 2,3), and technical terms (passive, fresh, fixed ...) are used before definitions, making the paper difficult to follow.
Questions for authors
  • What is the meaning of fixed and fresh source?
  • If increasing network capacity and training budget allows the model to recover performance without propagation, doesn't this rather suggest that propagation is not necessary? If so, is the main benefit of ASM limited to learning efficiency? Could other spatial mixing operators provide similar or even better efficiency?
1 comment

Thank you for the review. You worked through a paper that was hard to read and still asked specific questions, and we are grateful for that.

On the first question, a source is the random 4x4 noise that is given as input to the model. The fixed and fresh sources are two lists of 320 sources each. Both are random noise. We used the scores on the fixed list to pick the settings of the model and report the generation scores on the fresh list, so the reported scores had no part in those choices. The difference between the two lists is minor, so they did not need names, let alone a place in the abstract.

On the second question, the model without propagation does learn to produce the requested class once it is scaled up: with 2.7 times the parameters and 8.6 times the training of the paper's small model, it reaches 90.0% class accuracy across three seeds, while the model with ASM reaches 96.5%. The benefit of ASM is not only efficiency. The model without propagation generates less varied samples of each digit. It generates a close match, as judged by a classifier, for 41% of the 5,000 validation digits, while the model with ASM does so for 66%. One other untrained operator comes close. A random convolution reaches 77.0% in the small model, where ASM reaches 83.1%. In the scaled model the two reach about the same class accuracy, but the convolution generates a close match for 60% of the validation digits and ASM for 66%.

We are withdrawing the submission so that we can rewrite it.

Reviewer_jTd4OpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

I do not understand enough physics/optics to review this paper.

Critical strengths

N/A

Critical weaknesses

N/A

Questions for authors

N/A

Reviewer_56nCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is extremely unpolished. The main text delivers no technical information but meaningless engineering details.

Critical strengths

See above.

Critical weaknesses

See above.

Questions for authors

See above.

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

Related papers

Discussion (0)

Sign in to comment.