arxiv: 2604.27529 · v1 · submitted 2026-04-30 · 💻 cs.CV

Recognition: unknown

Adjoint Inversion Reveals Holographic Superposition and Destructive Interference in CNN Classifiers

Kaixiang Shu

Authors on Pith no claims yet

Pith reviewed 2026-05-07 08:00 UTC · model grok-4.3

classification 💻 cs.CV

keywords CNN interpretabilityadjoint inversiondestructive interferenceholographic superpositionspatial funnel hypothesischannel selectionout-of-distribution detection

0 comments

The pith

CNN classifiers operate by destructive interference that cancels a shared background direction to assemble class-specific residuals.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper challenges the assumption that convolutional networks first remove background pixels in their encoders before selecting from a clean set of features. It introduces an inversion method that avoids creating false signals in the output images. The reconstructions show that every channel produces the same visual result whether its weight is positive or negative. Only the algebraic combination of these identical-looking maps produces a sharp focus on the object. This pattern indicates that the network works by cancelling a common background pattern across the entire pixel space and keeping only the differences that distinguish the target class.

Core claim

Per-channel inversions in vision encoders are uniformly holographic, meaning positive and negative weight reconstructions are visually and energetically indistinguishable, yet their algebraic sum concentrates on the foreground. This establishes that classification proceeds via destructive interference, where classifier weights cancel a shared background direction in pixel space and constructively assemble class-discriminative residuals, directly falsifying the Spatial Funnel Hypothesis.

What carries the argument

The hallucination-free adjoint inversion framework that uses magnitude-phase decoupling and Local Adjoint Correctors to guarantee that every spatial gradient in a reconstruction originates only from genuinely active channels.

If this is right

The volume of the admissible interference subspace is the geometric quantity that determines how many channels are required for reliable classification.
This volume is mathematically dual to the GAP covariance determinant, which yields a covariance-volume channel selection algorithm carrying a (1-1/e) approximation guarantee.
Out-of-distribution failure appears as a measurable collapse of the covariance volume needed to sustain the interference process.
The same inversion framework applies without retraining to attention-based classifier heads.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same interference geometry might be testable in transformer models by applying the adjoint inversion to their attention weights.
Training procedures could be modified to enlarge the interference subspace volume explicitly, potentially improving robustness.
Interpretability methods that discard phase information are likely to miss the cancellation effects that actually drive decisions.

Load-bearing premise

That the magnitude-phase decoupling and Local Adjoint Correctors mathematically guarantee the spatial gradient support of every reconstruction stems strictly from genuinely active channels with no hallucinations.

What would settle it

An experiment in which the algebraic sum of positive and negative per-channel reconstructions fails to concentrate on the foreground object while the separate maps remain similar would falsify the destructive interference account.

Figures

Figures reproduced from arXiv: 2604.27529 by Kaixiang Shu.

**Figure 2.** Figure 2: Foreground energy proportion FG on (a) CUB-200 and (b) Pet, across four ImageNetpretrained encoders. 35–50%, and similar margins persist on CUB-200, Pets, and Dogs. Vanilla Gradient yields the worst scores (SSIM ≤ 6.1), confirming the dual pathologies of Section 2.1; optimization-based baselines remain in a low-fidelity tier, indicating that heuristic priors cannot structurally resolve the underlying alge… view at source ↗

**Figure 3.** Figure 3: (a) Per-channel inversions V˜L−1,i(X) at the deepest stage, grouped by forward energy rank: every channel recovers a complete rendering of the full scene, falsifying P1. (b) Stage-wise inversions Xˆ 0–3, the class-directional reconstruction Xˆ(c) , and its sign hemispheres Xˆ (c) ± : deep stages retain the full scene, and the two hemispheres are nearly indistinguishable, falsifying P2. Section 2.5, the pix… view at source ↗

**Figure 4.** Figure 4: (a,b,e,f) Pruning ratio vs accuracy on CIFAR-100 (in-distribution), CIFAR-100-C (corrup view at source ↗

**Figure 6.** Figure 6: Layer-wise (large) and per-token (3×3) attention visualizations across three hybrid attention blocks. softmax probability P(Birman) vs ablation fraction. At 70% ablation, Descending collapses to 0.61, vs 0.83 for Random and 0.88 for Ascending; Ascending tracks above Random throughout the mid-ablation regime, confirming that low-ECR channels are genuinely tail-irrelevant. However, beyond 80% ablation, Asce… view at source ↗

read the original abstract

A foundational assumption in CNN interpretability -- that deep encoders suppress background pixels while classifiers merely select from a cleaned feature pool (the Spatial Funnel Hypothesis) -- remains untested due to spatial hallucinations in existing visualization tools. We address this by introducing a hallucination-free inversion framework built on magnitude-phase decoupling and Local Adjoint Correctors. Our method mathematically guarantees that the spatial gradient support of every reconstruction stems strictly from genuinely active channels. Using this framework as a geometric probe, we uncover the first pixel-level evidence of strong superposition in vision encoders. We show that per-channel inversions are uniformly holographic: positive and negative weight reconstructions are visually and energetically indistinguishable. However, their algebraic sum sharply concentrates on the foreground. This proves classification operates via destructive interference -- classifier weights cancel a shared background direction in pixel space and constructively assemble class-discriminative residuals, directly falsifying the Spatial Funnel Hypothesis. This interference model identifies the volume of the admissible interference subspace as the geometric quantity governing channel requirements. We prove this volume is dual to the GAP covariance determinant, yielding a covariance-volume channel selection algorithm with a $(1-1/e)$ approximation guarantee. This algorithm mathematically reveals out-of-distribution (OOD) failure as a measurable collapse of the covariance volume essential for interference-based classification. Our framework extends seamlessly to attention-based heads without retraining.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The inversion framework is a solid technical idea but the destructive interference claim doesn't survive the nonlinear layers in real CNNs.

read the letter

The paper's core contribution is a new inversion method built on magnitude-phase decoupling plus Local Adjoint Correctors that aims to produce hallucination-free reconstructions from CNN channels. It then interprets the results as evidence that classification works through destructive interference: positive and negative weight reconstructions look the same, yet their sum concentrates on the foreground while canceling background. From this it derives a channel-selection algorithm based on the volume of an admissible interference subspace, which it claims is dual to the GAP covariance determinant and comes with a (1-1/e) guarantee. That algorithmic piece is the part that could actually be used by others if the underlying geometry holds.

Referee Report

3 major / 3 minor

Summary. The paper introduces a hallucination-free inversion framework for CNN classifiers based on magnitude-phase decoupling and Local Adjoint Correctors, which it claims mathematically guarantees that reconstruction gradients derive only from active channels. Using this as a probe, it reports that per-channel positive and negative weight inversions are holographically indistinguishable yet algebraically sum to foreground pixels, interpreted as evidence of destructive interference that cancels a shared background direction in pixel space. This is presented as direct falsification of the Spatial Funnel Hypothesis. The work further claims a duality between the volume of the admissible interference subspace and the GAP covariance determinant, yielding a covariance-volume channel selection algorithm with a (1-1/e) approximation guarantee, and attributes OOD failures to measurable collapse of this volume. The framework is stated to extend to attention heads without retraining.

Significance. If the inversion framework's claimed mathematical guarantees hold and are invariant to standard CNN nonlinearities, the work would offer a geometrically grounded alternative to existing visualization methods and provide the first pixel-level evidence of superposition-based classification mechanisms. The interference model and its link to channel selection and OOD detection could influence interpretability research by shifting focus from feature selection to explicit cancellation dynamics. The submodular channel-selection result, if novel rather than a re-derivation of known results, would add a practical tool with approximation guarantees.

major comments (3)

[Abstract; inversion framework section] Abstract and inversion framework description: the central claim that magnitude-phase decoupling plus Local Adjoint Correctors 'mathematically guarantee' that spatial gradient support stems strictly from active channels (with no hallucinations) is load-bearing for the destructive-interference interpretation and the falsification of the Spatial Funnel Hypothesis. Standard CNNs contain ReLUs, pooling, and batch-norm that break the linearity and phase assumptions required for exact adjoint guarantees; the manuscript must supply the explicit derivation showing how the correctors enforce the property globally rather than only locally or under unstated approximations.
[Abstract; channel selection algorithm section] Interference subspace volume and GAP covariance duality (claimed in the abstract): the reported duality appears to reduce to a mathematical identity by construction rather than an empirical or predictive relation. If the admissible interference subspace volume is defined in terms of the same covariance structure used for the determinant, the 'duality' does not constitute an independent geometric discovery but a tautology; this undermines the novelty of the covariance-volume channel selection algorithm and its claimed (1-1/e) guarantee.
[Channel selection algorithm section] Channel-selection algorithm and submodularity claim: the (1-1/e) approximation guarantee is stated to follow from the volume-covariance duality. If the underlying set function is a standard coverage or determinant-based submodular function already analyzed in the literature on submodular maximization, the result is not new; the manuscript should clarify the precise objective function and prove that the guarantee is not simply an application of the classic Nemhauser et al. result.

minor comments (3)

[Methods] Notation for 'admissible interference subspace' and 'Local Adjoint Correctors' is introduced without a clear definition or pseudocode; a formal definition and algorithmic listing would improve reproducibility.
[Abstract] The abstract asserts 'proofs' and 'mathematical guarantees' but the provided text contains no theorem statements, lemmas, or derivation steps; these should be added with numbered equations.
[Extension to attention heads] The extension to attention-based heads is stated to be seamless; a brief description of the required modifications (or lack thereof) would strengthen the claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below. Where the comments identify needs for additional derivation or clarification, we will revise the manuscript accordingly to strengthen the presentation without altering the core claims.

read point-by-point responses

Referee: [Abstract; inversion framework section] Abstract and inversion framework description: the central claim that magnitude-phase decoupling plus Local Adjoint Correctors 'mathematically guarantee' that spatial gradient support stems strictly from active channels (with no hallucinations) is load-bearing for the destructive-interference interpretation and the falsification of the Spatial Funnel Hypothesis. Standard CNNs contain ReLUs, pooling, and batch-norm that break the linearity and phase assumptions required for exact adjoint guarantees; the manuscript must supply the explicit derivation showing how the correctors enforce the property globally rather than only locally or under unstated approximations.

Authors: We agree that the nonlinearities present in standard CNNs (ReLUs, pooling, and batch-norm) require explicit handling to uphold the adjoint guarantees. The Local Adjoint Correctors are designed to compensate for these effects through local magnitude-phase adjustments that propagate to global consistency. In the revised manuscript we will add a self-contained derivation in the inversion framework section that walks through the compensation step by step, showing how the correctors restore the property that reconstruction gradients derive strictly from active channels even after the nonlinear operations. revision: yes
Referee: [Abstract; channel selection algorithm section] Interference subspace volume and GAP covariance duality (claimed in the abstract): the reported duality appears to reduce to a mathematical identity by construction rather than an empirical or predictive relation. If the admissible interference subspace volume is defined in terms of the same covariance structure used for the determinant, the 'duality' does not constitute an independent geometric discovery but a tautology; this undermines the novelty of the covariance-volume channel selection algorithm and its claimed (1-1/e) guarantee.

Authors: The admissible interference subspace is defined geometrically as the span of pixel-space directions that the classifier weights can cancel without changing the class logit; this definition does not presuppose the GAP covariance matrix. We then prove that the volume of this independently defined subspace equals the determinant of the GAP covariance. The relation is therefore a derived geometric identity rather than a definitional tautology. We will revise the abstract and algorithm section to state the two definitions separately before presenting the proof, thereby clarifying that the duality constitutes a non-trivial link between interference geometry and covariance structure. revision: partial
Referee: [Channel selection algorithm section] Channel-selection algorithm and submodularity claim: the (1-1/e) approximation guarantee is stated to follow from the volume-covariance duality. If the underlying set function is a standard coverage or determinant-based submodular function already analyzed in the literature on submodular maximization, the result is not new; the manuscript should clarify the precise objective function and prove that the guarantee is not simply an application of the classic Nemhauser et al. result.

Authors: The objective function maximized by the algorithm is the volume of the admissible interference subspace (equivalently, the log-determinant of the GAP covariance). While the (1-1/e) guarantee for monotone submodular functions is classical, the novelty lies in showing that this particular volume function—derived from the holographic interference model—is submodular and that its maximization directly controls classification robustness. In the revision we will state the objective function explicitly, supply a self-contained proof of its submodularity, and include a brief comparison with prior determinant-based submodular results to delineate the contribution. revision: yes

Circularity Check

3 steps flagged

Interference-subspace volume duality is a definitional identity; channel-selection algorithm with (1-1/e) guarantee re-expresses standard submodular optimization

specific steps

self definitional [Abstract, paragraph 3]
"This interference model identifies the volume of the admissible interference subspace as the geometric quantity governing channel requirements. We prove this volume is dual to the GAP covariance determinant, yielding a covariance-volume channel selection algorithm with a (1-1/e) approximation guarantee."

The volume is defined as the governing geometric quantity inside the interference model; the subsequent 'proof' that it is dual to the GAP covariance determinant is therefore an identity that holds by how the subspace was constructed from the covariance, not an independent mathematical result.
renaming known result [Abstract, paragraph 3]
"yielding a covariance-volume channel selection algorithm with a (1-1/e) approximation guarantee. This algorithm mathematically reveals out-of-distribution (OOD) failure as a measurable collapse of the covariance volume essential for interference-based classification."

The (1-1/e) approximation guarantee is the textbook bound for the greedy algorithm on monotone submodular set functions; the 'new' covariance-volume algorithm is a reparametrization of this standard result rather than a novel derivation specific to the holographic interference model.
self definitional [Abstract, paragraph 1]
"We address this by introducing a hallucination-free inversion framework built on magnitude-phase decoupling and Local Adjoint Correctors. Our method mathematically guarantees that the spatial gradient support of every reconstruction stems strictly from genuinely active channels."

The guarantee is asserted as a direct consequence of the framework's own construction (magnitude-phase decoupling plus the correctors); it is therefore true by design of the inversion procedure rather than a derived property that could falsify the Spatial Funnel Hypothesis independently of the method's assumptions.

full rationale

The paper's load-bearing geometric claims reduce to identities or known results by construction. The admissible interference subspace volume is introduced as the quantity governing channel requirements and then 'proved' dual to the GAP covariance determinant; this duality follows tautologically once the subspace is defined from the same covariance structure. The resulting selection algorithm inherits the (1-1/e) guarantee directly from the greedy algorithm for monotone submodular maximization, without an independent derivation. The hallucination-free guarantee of the adjoint inversion is likewise built into the Local Adjoint Correctors by design rather than derived from the CNN's actual nonlinear forward pass. These steps make the central 'proof' of destructive interference and the channel-selection contribution re-expressions of the framework's inputs rather than independent predictions.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claims rest on the correctness of the adjoint inversion framework and the existence of a duality between interference volume and GAP covariance; no explicit free parameters are named but the approximation algorithm implicitly relies on submodular properties.

axioms (2)

standard math Linear algebra operations on channel activations and classifier weights produce valid pixel-space reconstructions
Invoked when claiming holographic per-channel inversions and their algebraic sum
domain assumption The spatial gradient support of reconstructions stems strictly from genuinely active channels
Stated as a mathematical guarantee of the Local Adjoint Correctors

invented entities (1)

admissible interference subspace no independent evidence
purpose: Geometric quantity whose volume governs channel requirements for classification
Introduced to explain why certain channels are needed; no independent falsifiable handle provided in abstract

pith-pipeline@v0.9.0 · 5534 in / 1325 out tokens · 28083 ms · 2026-05-07T08:00:50.937182+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 5 canonical work pages · 2 internal anchors

[1]

Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks

Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In2018 IEEE winter conference on applications of computer vision (WACV), pages 839–847. IEEE, 2018

2018
[2]

On selecting a maximum volume sub-matrix of a matrix and related problems.Theoretical Computer Science, 410(47-49):4801–4811, 2009

Ali Civril and Malik Magdon-Ismail. On selecting a maximum volume sub-matrix of a matrix and related problems.Theoretical Computer Science, 410(47-49):4801–4811, 2009

2009
[3]

Inverting visual representations with convolutional networks

Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4829–4837, 2016

2016
[4]

Toy Models of Superposition

Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition.arXiv preprint arXiv:2209.10652, 2022

work page internal anchor Pith review arXiv 2022
[5]

Adversarial robustness as a prior for learned representations.arXiv preprint arXiv:1906.00945, 2019

Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Adversarial robustness as a prior for learned representations.arXiv preprint arXiv:1906.00945, 2019

work page arXiv 1906
[6]

Interpretable explanations of black boxes by meaningful perturbation

Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. InProceedings of the IEEE international conference on computer vision, pages 3429–3437, 2017

2017
[7]

A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.The journal of machine learning research, 13(1):723–773, 2012

2012
[8]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

2016
[9]

Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations.arXiv preprint arXiv:1903.12261, 2019

work page internal anchor Pith review arXiv 1903
[10]

Densely connected convolutional networks

Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

2017
[11]

Novel dataset for fine-grained image categorization: Stanford dogs

Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. InProc. CVPR workshop on fine-grained visual categorization (FGVC), volume 2, 2011

2011
[12]

Determinantal point processes for machine learning.Foundations and Trends® in Machine Learning, 5(2-3):123–286, 2012

Alex Kulesza and Ben Taskar. Determinantal point processes for machine learning.Foundations and Trends® in Machine Learning, 5(2-3):123–286, 2012

2012
[13]

A convnet for the 2020s

Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976–11986, 2022. 12

2022
[14]

Understanding deep image representations by invert- ing them

Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by invert- ing them. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5188–5196, 2015

2015
[15]

An analysis of approximations for maximizing submodular set functions—i.Mathematical programming, 14(1):265–294, 1978

George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. An analysis of approximations for maximizing submodular set functions—i.Mathematical programming, 14(1):265–294, 1978

1978
[16]

Plug & play generative networks: Conditional iterative generation of images in latent space

Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4467–4477, 2017

2017
[17]

Feature visualization.Distill, 2(11):e7, 2017

Chris Olah, Alexander Mordvintsev, and Ludwig Schubert. Feature visualization.Distill, 2(11):e7, 2017

2017
[18]

Cats and dogs

Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In2012 IEEE conference on computer vision and pattern recognition, pages 3498–3505. IEEE, 2012

2012
[19]

RISE: Randomized Input Sampling for Explanation of Black-box Models

Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models.arXiv preprint arXiv:1806.07421, 2018

work page Pith review arXiv 2018
[20]

Grad-cam: Visual explanations from deep networks via gradient-based localization

Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. InProceedings of the IEEE international conference on computer vision, pages 618–626, 2017

2017
[21]

The caltech-ucsd birds-200-2011 dataset

Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011

2011
[22]

nail bed

Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24–25, 2020. A Additional Method Details A.1 Detailed Diagnosis of the Two...

2020
[23]

activations move to a different part of feature space

By the chain rule,∂h l,c(r, s)/∂X(i, j) = 0. The summand is0·0 = 0. Sub-case B2: zl,c(r, s) = 0 .The ReLU is non-differentiable at zero. Following the standard convention adopted by all major automatic-differentiation frameworks (PyTorch, JAX, TensorFlow), the sub-gradient at z= 0 is set to zero, again yielding ∂hl,c(r, s)/∂X(i, j) = 0 via the chain rule,...

work page arXiv 2011