Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Hierarchical Feature-level Reverse Propagation for Post-Training Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A pretrained network can be post-trained by solving for the feature maps its frozen layers would need in order to classify correctly, replacing end-to-end backpropagation with systems of linear equations and least-squares problems.

desk verdict A target-propagation-style post-training scheme with a few clean closed-form derivations, but the rigorous claim rests on an unquantified approximation and the experiments do not support the advertised gains. read the letter →

arxiv 2506.07188 v1 pith:7GM2MPHI submitted 2025-06-08 cs.CV

Ni Ding , Lei He , Shengbo Eben Li , Keqiang Li This is my paper
classification cs.CV MSC 68T0790C2090C2565T50
keywords hierarchicaldecoupledpost-trainingfeaturemapreconstructionbackpropagationoptimalembeddingexplainableAIFFT-basedconvolutionimageclassificationautonomousdriving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a pretrained neural network can be improved module by module without end-to-end backpropagation: first compute, from the top layer downward, what the intermediate feature maps would have to be for the frozen upper layers to classify each image correctly, then train the lower modules to reproduce those reconstructed features. The reverse computation is formalized as convex optimization — a system of linear equations when the layer's feature dimension shrinks in the forward direction (the underdetermined case) and a least-squares problem when it grows (the overdetermined case) — with convolutions solved frequency-wise in the Fourier domain and the scalar label mapped to an output vector by a provably optimal embedding. If the claim holds, any contiguous block of a pretrained network can be post-trained in isolation using label-derived feature targets, and the deviation between forward and reconstructed features becomes a measurable per-layer diagnostic. The experiments on six image classification benchmarks support the claim for blocks near the output, where the reconstructed targets are reliable, and the paper attributes its weaker early-layer performance to accumulated reverse-computation error.

What carries the argument

Two principles carry the argument. The minimal deviation principle (MDP) applies when the reconstruction equation is underdetermined: among the infinitely many features that would reproduce the next layer's target exactly, choose the one closest to the original forward feature under the L2 norm; this is a strictly convex quadratic program whose KKT conditions collapse into the linear system of Eq. 6. The computing consistency principle (CCP) applies when the equation is overdetermined: choose the feature that minimizes the reconstruction residual, a linear least-squares problem (Eq. 7). For convolutional layers the same dichotomy is solved in the Fourier domain, where flipping the kernel and zero-padding turn the network convolution into a pointwise product, so each spatial frequency becomes an independent small linear system built from FFTs — this is how the paper avoids dense linear algebra on full feature maps. The optimal embedding closes the loop at the top: the scalar label is replaced by the projection of the forward output vector onto the set of vectors whose argmax is the label, with the L1 norm solved by maximum assignment (provably optimal) and the L2 norm by a KKT active-set algorithm (nearest embedding).

What would settle it

Reconstruct the feature maps for a trained network, then feed each reconstructed $a^*_l$ forward through the frozen layer it was solved against and compare with the target $z^*_{l+1}$: an exact MDP solution must match to numerical precision, so a residual that grows with depth directly measures the boundary-correction error. A reader could then run post-training twice, once with $G^x_{n,m}[\hat a_l]$ and once with the boundary term solved to a fixed point, and check whether the accuracy gap between FR-PT and backpropagation-based post-training survives the change.

Watch

Extended reading notes

Core claim

The central claim is that feature-level reverse computation — recovering the activations that would turn a desired label into the network's output — can be posed as well-posed optimization problems rather than heuristic inversion. At each linear layer the forward map $W_l a_l + b_l = z_{l+1}$ is either underdetermined (when $n_l \ge n_{l+1}$) or overdetermined (when $n_l < n_{l+1}$). In the underdetermined case the paper selects the feature minimizing $\|a_l - \hat a_l\|_2$ subject to exact reconstruction — the minimal deviation principle, whose KKT conditions collapse into a linear system; in the overdetermined case it selects the feature minimizing the reconstruction residual $\|W_l a_l + b_l - z^*_{l+1}\|_2$ — the computing consistency principle, a linear least-squares problem. For convolutional layers the same dichotomy is solved frequency-wise using the FFT and the convolution theorem, with a boundary-correction term evaluated at the forward feature rather than the unknown reconstructed one. At the output layer a scalar label is embedded as the closest vector, under the L1 or L2 norm, whose argmax equals the label — the maximum-assignment solution for L1 and a KKT active-set nearest-embedding solution for L2. Modules between layers $l_S$ and $l_R$ are then trained by ordinary backpropagation on a weighted sum of the classification loss and the deviation between forward and reconstructed features at layer $l_R$.

Load-bearing premise

In Section 3.2 the Fourier-domain reconstruction replaces the boundary-correction term $G^x_{n,m}[a^*_l]$ — which should be evaluated at the unknown reconstructed feature — with $G^x_{n,m}[\hat a_l]$, evaluated at the known forward feature, and the paper calls the resulting equation 'approximately correct' without bounding the error; because the reconstruction proceeds greedily layer by layer, that error is free to accumulate.

Editorial extensions

If this is right

  • Any contiguous block of a pretrained CNN can be post-trained in isolation: the label-derived reconstructed feature at the block's output supplies the supervision, so the rest of the network stays frozen during the entire update.
  • The reverse pass is deterministic and convex, so surrogate feature targets for an entire dataset can be precomputed once and reused across repeated post-training runs without re-entering the network.
  • On converged baselines, a single epoch of FR-PT can surpass ten epochs of backpropagation-based post-training, indicating that once a network has converged, the gains come from correcting the prediction process rather than from further iterative optimization.
  • Feature reconstruction succeeds where channel counts decrease in the forward direction (the exactly consistent MDP regime) and degrades where channels increase (the approximate CCP regime) or pooling discards information, which is why the experimental sweet spot is post-training blocks near the output.
  • The absolute difference between forward and reconstructed feature maps shrinks as a network's accuracy grows, giving a per-layer, label-informed visualization of training maturity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A likely transfer path the paper leaves untested: if the boundary-correction approximation in the Fourier-domain reconstruction can be made exact or bounded, the same MDP/CCP dichotomy should extend to ResNets and Transformers, where attention and normalization would be the genuinely hard operations because they have no natural "closest feature consistent with the next layer" analogue.
  • The greedy layer-by-layer scheme is the most plausible source of the early-layer accuracy loss; solving for an entire block of features jointly through one least-squares problem in the Fourier domain would directly test whether greedy choice, rather than the well-posed formulation, is the bottleneck.
  • A consequence the paper leaves implicit: the reconstruction pipeline doubles as a reachability test — if the optimal embedding demands an output vector far from the forward output, or if the deep reconstructed feature drifts outside what the frozen layers can produce, the network is being asked to represent a label its current trunk cannot express, and the deviation itself is a usable failure sign
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a hierarchical and decoupled post-training framework (FR-PT) for pretrained CNNs. Starting from a label, it first maps the label to a target output vector via an optimal embedding (L1 maximum assignment or L2 nearest embedding), then reconstructs intermediate feature maps backward through frozen layers by solving linear systems or least-squares problems for linear and convolutional operations, with pooling and nonlinear activations handled by heuristic reverse rules. The reconstructed feature at layer l_R is used together with the classification loss to post-train the module between l_S and l_R. The authors compare FR-PT with BP-based post-training on six image classification benchmarks and report accuracy and GPU memory, along with ablations of the embedding method and an analysis across training stages.

Significance. If the reconstruction step were genuinely well-posed and exact, the framework would be an interesting alternative to end-to-end backpropagation for modular post-training, and the feature-deviation visualization could contribute to interpretability. The paper contains some correct components: the L1 optimal-embedding proof, the L2 nearest-embedding derivation via KKT conditions, the linear-layer equality-constrained QP and least-squares formulations, and a reasonably large set of experiments. However, the central claim of a rigorous reformulation is undermined by the acknowledged approximation in the convolutional reconstruction and by the heuristic reverse operations for pooling and ReLU. Moreover, the experimental results do not consistently support the stated superiority in generalization or computational efficiency, since FR-PT is often within one standard deviation of BP-PT, worse on MNIST, and uses more GPU memory in every reported configuration. The contribution is therefore not established as presented.

major comments (4)
  1. [Section 3.2, Eq. (17)] The load-bearing approximation is stated in the text immediately after Eq. (17): in real computing, G^x_{n,m}[a*_l] is replaced by G^x_{n,m}[\hat a_l], making Eq. (17) only approximately correct. The unknown a*_l appears inside the boundary-correction term of the very constraint used to solve for a*_l, yet no Lipschitz estimate, error bound, or residual certificate is provided for this replacement. Because the reconstruction is greedy and layer-by-layer, the error can accumulate, so the final reconstructed feature a*_{l_R} may systematically deviate from any feature that the frozen subnetwork can actually produce. This directly contradicts the abstract's claim that feature-level reverse computation is 'rigorously reformulated' as well-posed linear systems. The authors should at minimum report consistency residuals ||z*_{l+1} - (conv(a*_l, K) + b_l)|| on the actual data, and should provide either a provable bound or an iterative correction scheme.
  2. [Section 3, full-rank assumption and Eqs. (6), (19), (20)] The manuscript assumes without validation that 'the linear weight matrix W^l is always full-rank' and relies on this for uniqueness in Eq. (6) and for the block systems in Eqs. (19) and (20). The assumption is not checked for the actual networks. For linear layers, the required condition is full row rank in the case n_l >= n_{l+1} and full column rank in the case n_l < n_{l+1}; for convolutional layers, the frequency-domain matrix F(\tilde K) can be rank-deficient (for example with symmetric kernels or certain channel counts), in which case the KKT matrix is singular and the claimed unique solution does not exist. The authors should report numerical ranks of the relevant matrices for the architectures used, or restrict the method to provably full-rank cases.
  3. [Tables 1-6] The experimental evidence does not support the abstract's claims of 'superior generalization performance and computational efficiency.' In many configurations the difference between FR-PT and BP-PT is within one standard deviation; on MNIST (Table 1) FR-PT is consistently below BP-PT; and GPU memory usage is higher for FR-PT in every listed configuration, with no wall-clock time reported. The paper should provide paired significance tests across the 10 runs, report a proper efficiency metric (e.g., end-to-end time and peak memory), and either demonstrate consistent statistically significant gains or substantially weaken the superiority claims.
  4. [Section 3.4] The reverse computation for nonlinear activations and pooling is heuristic: ReLU is inverted by the identity map, and pooling layers are reconstructed by directly copying values. These steps are not posed as the well-posed optimization problems that the paper claims as its main contribution, and their approximation error is not analyzed. Since these heuristic inverses are used in the same greedy backward chain as the linear and convolutional reconstructions, the overall reconstructed feature is not the solution of a single well-posed problem. The authors should either incorporate these operations into the optimization framework with explicit error control or explicitly limit the rigor claim to the linear and convolutional subproblems.
minor comments (5)
  1. [Section 3.1, Eq. (6) and surrounding text] In the Lagrangian definition, the term '\hat a_l W^l' appears to be a typo for 'W^l a_l', and the multiplier term should be written with a transpose; the subsequent KKT matrix in Eq. (6) is consistent with the corrected form.
  2. [Section 3.2, Eqs. (12)-(19)] The notation F(\tilde K) and F(a_l) in Eqs. (19) and (20) is ambiguous because these objects are not defined for multi-channel kernels and feature maps; the authors should define the matrix with entries F(\tilde K[n,m]) and the column vector with entries F(a_l[x,m]) at each frequency (u,v).
  3. [Section 3.2, Eq. (14)] The symbol 'T b,n_m' in Eq. (14) is not defined; it should presumably be T^x_{n,m} after zero-padding.
  4. [Tables 1-6] The loss coefficient alpha is set per dataset with values 0.1, 0.3, or 0.7, but no sensitivity analysis or selection criterion is provided; the reported conclusions depend on this free parameter.
  5. [Section 1, Contributions] The claim that this is 'the first method to formalize feature-level reverse computation as well-posed optimization problems' is presented without discussing prior reverse-computation and synthetic-target methods in the related work; the wording should be tempered or supported by a comparative discussion.

Circularity Check

0 steps flagged · score 0.0 of 10

No equation-level circularity: the reconstructed feature targets are self-generated but are not fitted to the reported accuracies, and the G-substitution in Eq. 17 is an approximation issue rather than a tautological reduction.

full rationale

Walking the derivation chain, the reconstructed target a*_L is defined as the nearest output vector to the forward logits whose argmax equals the label (Eqs. 21-25), and each lower-layer feature a*_l is defined as a least-norm or least-squares preimage of z*_{l+1} under the frozen weight map (Eqs. 4, 6, 7, 18-20). These targets are self-generated in the sense that they are computed from the same frozen network's forward features, weights, and the external label, but they are not fitted to the test accuracies that the paper reports. The only per-dataset tuning parameter is the scalar alpha in Eq. 3, and the headline comparison is against external test accuracy, so the reported superiority claims are not forced by construction. The convolutional derivation does contain the explicit approximation after Eq. 17: 'In real computing, the G^x_{n,m}[a*_l] is replaced by G^x_{n,m}[hat a_l], leading formula 17 to be approximately correct.' That substitution makes the FFT-based solve inexact and could accumulate error through the greedy layer-by-layer reconstruction, but it does not equate the solver's output to its input by definition; the constraint remains anchored by the reconstructed z*_{l+1} and the frozen weights. The full-rank assumption on W_l is asserted without validation, which is a rigor concern, not a circular one. The only author-overlapping citation in the manuscript is about feature-map convergence in related work and is not load-bearing for the central derivation. Overall, no step reduces a claimed prediction or first-principles result to its own inputs, so the circularity score is 0.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central method rests on the frozen weights being invertible in a well-posed sense, the full-rank assumption, on a greedy reconstruction that ignores downstream consistency, and on an acknowledged boundary-term approximation in the FFT-based convolution inversion. The only per-dataset free parameter is the loss coefficient alpha. No new physical entities are introduced.

free parameters (1)
  • loss coefficient alpha = 0.1 (MNIST, CIFAR-10, ImageWoof), 0.3 (ImageNette), 0.7 (CIFAR-100, Tiny ImageNet)
    Chosen per dataset in the reported experiments; no sensitivity analysis is provided, and the comparison between FR-PT and BP-PT may change with alpha.
assumptions (5)
  • ad hoc to paper All linear weight matrices W^l are full-rank
    Stated in Section 3 before Eq. 6; guarantees uniqueness of the KKT system. Trained CNNs are not guaranteed to have full-rank weights, and the paper does not check this on its baselines.
  • domain assumption Greedy layer-by-layer reconstruction is adequate
    Stated in Section 3: each reconstruction step ignores whether the chosen a*_l is optimal for later, earlier, reconstructions. Accumulated suboptimality is not analyzed.
  • ad hoc to paper Boundary correction G can be evaluated at forward features
    Section 3.2, text after Eq. 17: G^x_{n,m}[a*_l] is replaced by G^x_{n,m}[\hat a_l], making the convolutional reconstruction only approximately consistent with the frozen layer.
  • domain assumption ReLU reverse is identity and pooling reverse copies values
    Section 3.4: non-bijective activations such as ReLU are inverted by identity, and max-pooling is inverted by copying post-pooling values, losing information that is asserted to be insignificant.
  • standard math Fourier convolution theorem and isometry
    Used in Section 3.2 to diagonalize convolutions; standard and correctly applied modulo boundary corrections.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Feature-level Reverse Propagation for Post-Training Neural Networks." pith.science (2026). https://pith.science/paper/7GM2MPHI

@misc{pith2026250607188,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Feature-level Reverse Propagation for Post-Training Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7GM2MPHI}},
  note         = {Machine review of arXiv:2506.07188}
}
read the original abstract

End-to-end autonomous driving has emerged as a dominant paradigm, yet its highly entangled black-box models pose significant challenges in terms of interpretability and safety assurance. To improve model transparency and training flexibility, this paper proposes a hierarchical and decoupled post-training framework tailored for pretrained neural networks. By reconstructing intermediate feature maps from ground-truth labels, surrogate supervisory signals are introduced at transitional layers to enable independent training of specific components, thereby avoiding the complexity and coupling of conventional end-to-end backpropagation and providing interpretable insights into networks' internal mechanisms. To the best of our knowledge, this is the first method to formalize feature-level reverse computation as well-posed optimization problems, which we rigorously reformulate as systems of linear equations or least squares problems. This establishes a novel and efficient training paradigm that extends gradient backpropagation to feature backpropagation. Extensive experiments on multiple standard image classification benchmarks demonstrate that the proposed method achieves superior generalization performance and computational efficiency compared to traditional training approaches, validating its effectiveness and potential.

Figures

Figures reproduced from arXiv: 2506.07188 by the authors.

Figure 1
Figure 1. (a) Conventional back propagation training approach on CNN for image classification tasks. (b) Our hierarchical and decoupled post-training framework based on feature map reconstruction for image classification tasks. On the contrary, the modularization paradigm comprises a sequence of refined components for distinct subtasks such as object detection, trajectory prediction, and route planning. Modular architectures … view at source ↗
Figure 2
Figure 2. Hierarchical and Decoupled Post-Training based on Feature Map Reconstruction for Image Classification Tasks. Given a well-trained CNN, the optimal embedding and feature reconstruction are executed in line with the network’s backward computation flow to obtain the reconstructed (lR) th feature map. A combination of reconstruction loss and classification loss is then used to post-train the intermediate modules from l … view at source ↗
Figure 3
Figure 3. Schematic diagram of the boundary set B0,1 for Ke (0, 1). F(T x n,m)(u, v) = H Xl+1 k=1 W Xl+1 j=1 T b,n m (k, j)e −2πi( u(k+HK−1) Hl + v(j+WK−1) Wl ) = H Xl+1 k=1 W Xl+1 j=1 X HK s=1 X WK t=1 aˆl [x, m](k+HK−1−s,j+WK−1−t)· Ke[n, m](s, t)e −2πi( u(k+HK−1) Hl + v(j+WK−1) Wl ) (14) Due to the finiteness of feature map size, the boundary of feature maps needs further modification using Eq. 15, where set Bs,t for each (… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Absolute values of discrepancy between forward feature maps [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Average test accuracy with 95% confidence of optimal embedding approaches compared with one-hot coding across six benchmarks. underfit convergence overfit [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: The post-training results comparison across 30 baselines with different training stages. A single baseline for Cifar100 evolves by one BP epoch per stage (black line). Then, for each post-training, only the parameters between 1 th layer and 4 th layer are trainable. co…
Figure 7
Figure 7. Figure 7: Influence comparison of FR-PT on different ”3Conv+2fc” baselines of cifar10. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    Towards explainable semantic segmentation for autonomous driving systems by multi-scale variational attention

    Mohanad Abukmeil, Angelo Genovese, Vincenzo Piuri, Francesco Rundo, and Fabio Scotti. Towards explainable semantic segmentation for autonomous driving systems by multi-scale variational attention. In2021 IEEE In- ternational Conference on Autonomous Systems (ICAS), pages 1–5, 2021

    work page 2021

  2. [2]

    Mayank Bansal, Alex Krizhevsky, and Abhijit S. Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst.ArXiv, abs/1812.03079, 2018

    arXiv 2018

  3. [3]

    Decoupled greedy learning of cnns

    Eugene Belilovsky, Michael Eickenberg, and Edouard Oyallon. Decoupled greedy learning of cnns. InInterna- tional Conference on Machine Learning, pages 736–745. PMLR, 2020

    work page 2020

  4. [4]

    Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba

    Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard Firner, Beat Flepp, Pra- soon Goyal, Lawrence D. Jackel, Mathew Monfort, Urs Muller, Jiakai Zhang, Xin Zhang, Jake Zhao, and Karol Zieba. End to end learning for self-driving cars, 2016

    work page 2016

  5. [5]

    Cambridge University Press, 2004

    Stephen Boyd and Lieven Vandenberghe.Convex Opti- mization. Cambridge University Press, 2004

    work page 2004

  6. [6]

    Nadia Burkart and Marco F. Huber. A survey on the explainability of supervised machine learning.J. Artif. Int. Res., 70:245–317, May 2021

    work page 2021

  7. [7]

    Vatld: A visual analytics system to assess, understand and im- prove traffic light detection.IEEE Transactions on Visu- alization and Computer Graphics, 27(2):261–271, 2021

    Liang Gou, Lincan Zou, Nanxiang Li, Michael Hofmann, Arvind Kumar Shekar, Axel Wendt, and Liu Ren. Vatld: A visual analytics system to assess, understand and im- prove traffic light detection.IEEE Transactions on Visu- alization and Computer Graphics, 27(2):261–271, 2021

    work page 2021

  8. [8]

    Emma: End-to- end multimodal model for autonomous driving, 2024

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to- end multimodal model for autonomous driving, 2024

    work page 2024

Show all 40 references
  1. [9]

    Decoupled neural interfaces using synthetic gradients, 2017

    Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David Silver, and Koray Kavukcuoglu. Decoupled neural interfaces using synthetic gradients, 2017

    2017

  2. [10]

    Layercam: Exploring hierarchical class activation maps for localization.IEEE Transactions on Image Processing, 30:5875–5888, 2021

    Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming- Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization.IEEE Transactions on Image Processing, 30:5875–5888, 2021

    2021

  3. [11]

    Intention-aware interactive transformer for real-time ve- hicle trajectory prediction in dense traffic.Transportation Research Record, 2677(3):946–960, 2023

    Titong Jiang, Yahui Liu, Qing Dong, and Tao Xu. Intention-aware interactive transformer for real-time ve- hicle trajectory prediction in dense traffic.Transportation Research Record, 2677(3):946–960, 2023

    2023

  4. [12]

    Diffstack: A differentiable and modular control stack for autonomous vehicles

    Peter Karkus, Boris Ivanovic, Shie Mannor, and Marco Pavone. Diffstack: A differentiable and modular control stack for autonomous vehicles. In Karen Liu, Dana Kulic, and Jeff Ichnowski, editors,Proceedings of The 6th Con- ference on Robot Learning, volume 205 ofProceedings of ...

    2023

  5. [13]

    Boosting monocular 3d object detection with object-centric auxiliary depth supervision

    Youngseok Kim, Sanmin Kim, Sangmin Sim, Jun Won Choi, and Dongsuk Kum. Boosting monocular 3d object detection with object-centric auxiliary depth supervision. IEEE Transactions on Intelligent Transportation Systems, 24(2):1801–1813, 2023

    2023

  6. [14]

    Explainable action prediction through self-supervision on scene graphs

    Pawit Kochakarn, Daniele De Martini, Daniel Omeiza, and Lars Kunze. Explainable action prediction through self-supervision on scene graphs. In2023 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pages 1479–1485, 2023

    2023

  7. [15]

    Suresh Kolekar, Shilpa Gite, Biswajeet Pradhan, and Ab- dullah Alamri. Explainable ai in scene understanding for autonomous vehicles in unstructured traffic environ- ments on indian roads using the inception u-net model with grad-cam visualization.Sensors, 22(24), 2022

    2022

  8. [16]

    Albrecht

    Anton Kuznietsov, Balint Gyevnar, Cheng Wang, Steven Peters, and Stefano V. Albrecht. Explainable ai for safe and trustworthy autonomous driving: A systematic re- view.IEEE Transactions on Intelligent Transportation Systems, 25(12):19342–19364, 2024

    2024

  9. [17]

    Exploring intermediate representation for monocular vehicle pose estimation

    Shichao Li, Zengqiang Yan, Hongyang Li, and Kwang- Ting Cheng. Exploring intermediate representation for monocular vehicle pose estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1873–1883, June 2021

    2021

  10. [18]

    Od-xai: Explainable ai-based semantic object detection for au- tonomous vehicles.Applied Sciences, 12(11), 2022

    Harsh Mankodiya, Dhairya Jadav, Rajesh Gupta, Sudeep Tanwar, Wei-Chiang Hong, and Ravi Sharma. Od-xai: Explainable ai-based semantic object detection for au- tonomous vehicles.Applied Sciences, 12(11), 2022

    2022

  11. [19]

    Deep learning techniques: an overview.Advanced Machine Learning Technologies and Applications: Proceedings of AMLTA 2020, pages 599–608, 2021

    Amitha Mathew, P Amudha, and S Sivakumari. Deep learning techniques: an overview.Advanced Machine Learning Technologies and Applications: Proceedings of AMLTA 2020, pages 599–608, 2021

    2020

  12. [20]

    Layerwise knowledge extraction from deep convolutional networks

    Simon Odense and Artur d’ Avila Garcez. Layerwise knowledge extraction from deep convolutional networks. arXiv preprint arXiv:2003.09000, 2020. 13

    2003 arXiv

  13. [21]

    Clip-bevformer: Enhancing multi-view image-based bev detector with ground truth flow, 2024

    Chenbin Pan, Burhaneddin Yaman, Senem Velipasalar, and Liu Ren. Clip-bevformer: Enhancing multi-view image-based bev detector with ground truth flow, 2024

    2024

  14. [22]

    Decoupled neural network training with re-computation and weight predic- tion.PloS one, 18(2):e0276427, 2023

    Jiawei Peng, Yicheng Xu, Zhiping Lin, Zhenyu Weng, Zishuo Yang, and Huiping Zhuang. Decoupled neural network training with re-computation and weight predic- tion.PloS one, 18(2):e0276427, 2023

    2023

  15. [23]

    Roth, and Horst Bischof

    David Schinagl, Georg Krispel, Horst Possegger, Peter M. Roth, and Horst Bischof. Occam’s laser: Occlusion-based attribution maps for 3d object detectors on lidar data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1141– 115...

    2022

  16. [24]

    Roth, and Horst Bischof

    David Schinagl, Georg Krispel, Horst Possegger, Peter M. Roth, and Horst Bischof. Occam’s laser: Occlusion-based attribution maps for 3d object detectors on lidar data. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1131–1140, 2022

    2022

  17. [25]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep net- works via gradient-based localization. In2017 IEEE International Conference on Computer Vision (ICCV), pages 618–626, 2017

    2017

  18. [26]

    Deep inside convolutional networks: Visualising image classification models and saliency maps.CoRR, abs/1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisser- man. Deep inside convolutional networks: Visualising image classification models and saliency maps.CoRR, abs/1312.6034, 2013

    2013 arXiv

  19. [27]

    A survey of end-to-end driving: Architectures and training meth- ods.IEEE Transactions on Neural Networks and Learning Systems, 33(4):1364–1384, 2022

    Ardi Tampuu, Tambet Matiisen, Maksym Semikin, Dmytro Fishman, and Naveed Muhammad. A survey of end-to-end driving: Architectures and training meth- ods.IEEE Transactions on Neural Networks and Learning Systems, 33(4):1364–1384, 2022

    2022

  20. [28]

    Lane-change inten- tion prediction of surrounding vehicles using bilstm-crf models with rule embedding

    Kai Wang, Jie Hou, and Xianlin Zeng. Lane-change inten- tion prediction of surrounding vehicles using bilstm-crf models with rule embedding. In2022 China Automation Congress (CAC), pages 2764–2769, 2022

    2022

  21. [29]

    Dataset distillation with neural characteristic function: A minmax perspective, 2025

    Shaobo Wang, Yicun Yang, Zhiyuan Liu, Chenghao Sun, Xuming Hu, Conghui He, and Linfeng Zhang. Dataset distillation with neural characteristic function: A minmax perspective, 2025

    2025

  22. [30]

    Wong, Zhenguo Li, and Hengshuang Zhao

    Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Letters, 9(10):8186–8193, 2024

    2024

  23. [31]

    Bev- former v2: Adapting modern image backbones to bird’s- eye-view recognition via perspective supervision

    Chenyu Yang, Yuntao Chen, Hao Tian, Chenxin Tao, Xizhou Zhu, Zhaoxiang Zhang, Gao Huang, Hongyang Li, Yu Qiao, Lewei Lu, Jie Zhou, and Jifeng Dai. Bev- former v2: Adapting modern image backbones to bird’s- eye-view recognition via perspective supervision. In Proceedings of the...

    2023

  24. [32]

    Hakan Yekta Yatbaz, Mehrdad Dianati, and Roger Wood- man. Introspection of dnn-based perception functions in automated driving systems: State-of-the-art and open research challenges.IEEE Transactions on Intelligent Transportation Systems, 25(2):1112–1130, 2024

    2024

  25. [33]

    Sun-glare region recognition using visual explanations for traffic light detection

    Keisuke Yoneda, Naoki Ichihara, Hotsuyuki Kawanishi, Tadashi Okuno, Lu Cao, and Naoki Suganuma. Sun-glare region recognition using visual explanations for traffic light detection. In2021 IEEE Intelligent Vehicles Sym- posium (IV), pages 1464–1469, 2021

    2021

  26. [34]

    Understanding neural networks through deep visualization, 2015

    Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization, 2015

    2015

  27. [35]

    Visualizing and un- derstanding convolutional networks, 2013

    Matthew D Zeiler and Rob Fergus. Visualizing and un- derstanding convolutional networks, 2013

    2013

  28. [36]

    Fea- ture map convergence evaluation for functional module, 2024

    Ludan Zhang, Chaoyi Chen, Lei He, and Keqiang Li. Fea- ture map convergence evaluation for functional module, 2024

    2024

  29. [37]

    Shap- cam: Visual explanations for convolutional neural net- works based on shapley value

    Quan Zheng, Ziwei Wang, Jie Zhou, and Jiwen Lu. Shap- cam: Visual explanations for convolutional neural net- works based on shapley value. InEuropean conference on computer vision, pages 459–474. Springer, 2022

    2022

  30. [38]

    Genad: Generative end-to-end autonomous driving

    Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. InEuropean Conference on Com- puter Vision, pages 87–104. Springer, 2024

    2024

  31. [39]

    Learning deep features for dis- criminative localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for dis- criminative localization. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921–2929, 2016

    2016

  32. [40]

    Fully decoupled neural network learning using delayed gradients.IEEE transactions on neural networks and learning systems, 33(10):6013–6020, 2021

    Huiping Zhuang, Yi Wang, Qinglai Liu, and Zhiping Lin. Fully decoupled neural network learning using delayed gradients.IEEE transactions on neural networks and learning systems, 33(10):6013–6020, 2021

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.