Pith. sign in

REVIEW 4 major objections 6 minor 67 references

THIRDEYE: Cue-Aware Monocular Depth Estimation via Brain-Inspired Multi-Stage Fusion

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ThirdEye contends that feeding explicit edge, normal, and layout cues from frozen specialists into a V1-to-V3 memory-augmented fusion hierarchy is the right way to build accurate, interpretable monocular depth estimation with only modest…

desk verdict An unvalidated architecture proposal that could be useful if the claimed 3.1% memory ablation survives contact with experiments; worth a desk-reject with an invitation to return with real results. read the letter →

arxiv 2506.20877 v1 pith:LM3LPFFO submitted 2025-06-25 cs.CV cs.AI

Calin Teodor Ioan This is my paper
classification cs.CVcs.AI
keywords monoculardepthestimationmulti-cuefusionfrozenpre-trainedspecialistskey-valueworkingmemoryuncertaintygatingbiologicallyinspiredvisioncorticalhierarchy
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

ThirdEye's proposal is that monocular depth estimation should not trust a single network to rediscover visual cues on its own. Instead, it 'spoon-feeds' the image to three frozen specialist networks—HED for occlusion edges, SDPS-Net for surface normals, and HorizonNet for room layout—and fuses their outputs through a three-stage cortical-style hierarchy (V1→V2→V3) with a 32-slot key-value working-memory bank. The central claim is that this division of labour separates where cues come from from how they are combined, letting the model inherit the specialists' external supervision, require only modest fine-tuning, and produce high-resolution disparity maps that are accurate and interpretable. The paper details the architecture, the neuroscience analogies, and a full experimental protocol; the quantitative results are deferred to a future revision. The only reported number is a preliminary ~3.1% drop in the $\delta_1$ metric when the memory bank is disabled.

What carries the argument

The load-bearing machinery is the frozen cue-fusion stack: three off-the-shelf specialists emit edge, normal, and layout maps together with log-variance maps; each cue is scaled as $\tilde{C}=C\odot\exp(-\sigma)$ before entering a cortical-style hierarchy. The hierarchy is a V1→V2→V3 analog—a convolutional stem with oriented filters, a depth-wise separable integration layer with self-attention, and two Swin-Tiny context blocks—paired with a 32-slot key-value memory $M \in \mathbb{R}^{S \times D}$ that is read by cross-attention and updated by a gated write with a learnable decay rate. An adaptive-bins transformer head converts the fused representation to $K=64$ bin centres and per-pixel soft assignments, and a guided filter uses the edge map to upsample the resulting low-resolution disparity to full resolution. This machinery splits cue acquisition (frozen, externally supervised) from cue reasoning (learned fusion), which is what lets the model promise plug-and-play upgrades and interpretable depth estimates.

What would settle it

Train the full ThirdEye pipeline on KITTI with the specified loss and compare Abs-Rel and $\delta_1$ against a variant with all reliability gates fixed to zero and a no-memory variant: the central claim fails if the full model does not beat both under matched training, because the paper's own planned uncertainty-gating-off ablation would then show the core mechanism contributes nothing.

Watch

Extended reading notes

Core claim

ThirdEye contends that supplying explicit monocular cues—occlusion edges, surface normals, and global layout—from frozen pre-trained experts, and fusing them through a V1-style convolutional stem, a V2 integration layer, and a V3 context layer with a key-value memory bank, is a better path to monocular depth than end-to-end implicit regression. Each cue is gated by an exponential reliability weight $\tilde{C}=C\odot\exp(-\sigma)$ that down-weights untrustworthy regions before fusion; the memory bank is read by cross-attention and written with a learnable decay gate, forming the paper's analogue of persistent activity in early visual cortex. Decoding uses an adaptive-bins transformer head with $K=64$ bin centres and edge-guided upsampling, trained by a composite loss of scale-invariant, gradient, SSIM, and cue-weight terms. The paper claims this yields high-resolution disparity with minimal fine-tuning and reports a preliminary ~3.1% drop in $\delta_1$ when memory is disabled, with full quantitative results planned for a later revision.

Load-bearing premise

The central bet is that each frozen cue expert can supply a trustworthy reliability map from the same depth supervision used to train the model; the paper does not specify how those variance maps are learned or computed.

Editorial extensions

If this is right

  • Replacing a cue expert with a better or domain-specific one becomes a drop-in change: the fusion core and memory are untouched, so domain adaptation reduces to swapping a frozen specialist.
  • Unreliable cue regions are automatically down-weighted by reliability gating, giving the model an explicit mechanism for ignoring noisy normals on shiny surfaces instead of encoding their errors into the depth map.
  • The memory bank gives the network a temporal feedback loop, so sequential frames can consolidate consistent cues and forget conflicting ones without deepening the network.
  • A depth failure can be traced to a specific specialist or memory slot, since each cue pathway is externally supervised and modular.
  • Training the fusion stack with frozen specialists is light enough to fit in under 6 GB VRAM, which would make the approach practical for embedded or low-budget setups.

Reading between the lines

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

  • If the reliability-gating design is doing the work, the variance heads are the component to stress-test first; a natural extension is to train them with auxiliary supervision (e.g., cross-view photometric consistency) rather than the depth-only loss specified in the paper.
  • The memory bank's gated write is effectively a continual-learning rule, so an experiment beyond the paper would feed a KITTI-trained model a stream of indoor frames and measure whether slot consolidation prevents catastrophic forgetting without gradient updates.
  • The frozen-specialist-plus-fusion recipe is not intrinsically depth-specific; swapping the adaptive-bins head for a semantic-segmentation or normal-estimation decoder would test whether the V1→V2→V3 stack is a general dense-prediction front end.
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 / 6 minor

Summary. The paper proposes ThirdEye, a monocular depth estimation architecture that fuses frozen off-the-shelf cue extractors (HED, SDPS-Net, HorizonNet) through a V1–V2–V3 cortical hierarchy with a key–value working-memory bank, an adaptive-bins transformer head, and edge-guided upsampling. The manuscript provides architectural details, neuroscientific analogies, a composite training objective, memory read–write equations, and an extensive planned experimental protocol. No experimental results are reported: the abstract explicitly states that quantitative results will appear in a future revision, Section 7 contains only planned experiments, and the only numeric claim in the paper, a ~3.1% delta-1 drop when memory is disabled, is cited to an ablation study that does not appear anywhere in the manuscript.

Significance. The modular cue-aware design and the attempt to ground depth-estimation components in cortical working-memory mechanisms are conceptually interesting. If validated, the frozen-specialist pipeline could offer interpretability, plug-and-play upgrades, and a principled way to weigh cues by estimated reliability. The paper also proposes a concrete multi-benchmark protocol (KITTI, DDAD, NYU-v2, MegaDepth, TartanAir) and a set of biological-plausibility checks. However, none of these strengths is backed by evidence in the current manuscript. As an empirical contribution, the paper is incomplete: there are no accuracy numbers, no comparisons to baselines, no ablation tables, and no figures showing qualitative outputs. The significance of the architectural claims therefore cannot be assessed at this stage.

major comments (4)
  1. [Abstract and Section 7] The central claim that ThirdEye 'produces a high-resolution disparity map' is unsupported. The abstract explicitly states that 'quantitative results will appear in a future revision,' and Section 7 ('Planned Experiments') contains only dataset descriptions, metric definitions, training details, and expected outcomes, with no results, tables, or metric values for KITTI, NYU-v2, DDAD, or MegaDepth. Without any accuracy evidence, the core capability claim cannot be evaluated; this is the load-bearing omission of the manuscript.
  2. [Section 8.2] The only quantitative statement in the paper, a '~3.1% drop in δ1 when memory is disabled,' is attributed to 'Preliminary ablations (Sec. 4).' Section 4 contains no ablation study, no table, no dataset split, no training protocol, and no metric values. The citation is unverifiable and the claim cannot be checked against any presented data.
  3. [Sections 4.4 and 5.3] The memory mechanism is specified inconsistently. Algorithm 1 initializes M <- 0_{S×D} and passes it to fV3 without any update, while Algorithm 2 reads from an uninitialized M0 and performs a loop over V2_BLOCK that does not appear in Algorithm 1. Moreover, Eq. (10) defines the update as M_{t+1} = (1−η)M_t + η M̃ with η = σ(W_η[mean(X_{t+1})]), whereas Eq. (11) defines M_{t+1} = (1−η)M_t + η σ(QK^T/√d)V. These are different update rules, and neither specifies how M̃ or the top-k aggregated latent is computed. The memory contribution is therefore not fully specified.
  4. [Sections 4.1 and 5.2] The uncertainty maps σ_E,N,P are load-bearing: Eq. (1) and Eq. (4) scale every cue by exp(−σ), but no training objective or supervision for the variance heads is specified. The loss Lcue in Eq. (7) penalizes fusion weights w_C, not the variance heads, and the cue specialists are frozen. As written, the reliability gates could be arbitrary constants, and the Bayesian cue-integration interpretation is not supported by any training mechanism.
minor comments (6)
  1. [Section 4.5] The first sentence of the 'Adaptive-Bins Depth Decoder' subsection is incomplete: '(0.9 M parameters) but drop the final refinement stage' has no grammatical subject or main verb.
  2. [Section 5.1, Eq. (4)] Equation (4) contains an empty citation: 'mirroring Bayesian precision weighting [].' This should either be completed or removed.
  3. [References] The reference list contains duplicates and likely inconsistencies: [11] repeats [1], [13] and [36] are both the AdaBins paper, [28] repeats [6], [30] and [39] both cite the same RAM paper, and [31] and [40] both cite predictive coding networks. These should be consolidated or corrected.
  4. [Algorithm 2] Algorithm 2 uses M0 before it is initialized and does not match Algorithm 1's single V2 pass; the pseudocode should be reconciled with the architecture described in Section 3.
  5. [Section 7.1] The description of KITTI as '93 training stereo pairs' is misleading if this refers to scenes rather than image pairs; the dataset statistics should be clarified to match the Eigen split convention used in the paper.
  6. [Figure 1] Figure 1 is referenced in Section 3 and Section 8.1, but no actual diagram appears in the manuscript text; the pipeline overview should be included or the reference removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: ThirdEye is an unsupported proposal whose claimed results are deferred, not derived from its own inputs.

full rationale

The paper does not exhibit any derivation chain that reduces to its own inputs. It is an architecture proposal with no completed experiments: the abstract states that 'quantitative results will appear in a future revision,' and Section 7 describes only planned experiments. No equation is defined in terms of the quantity it is claimed to predict, no fitted parameter is renamed as a prediction, and there are no self-citations or author-imported uniqueness theorems that carry the argument. The uncertainty-gating mechanism in Eq. (1) is an architectural choice whose variance heads are under-specified, but under-specification is a completeness problem, not circularity. The one numeric claim in Section 8.2, a '~3.1% drop in δ1 when memory is disabled,' cites 'Preliminary ablations (Sec. 4),' but Section 4 contains no ablation table or experimental results; this is an unsupported, self-referential assertion about the authors' own model, and it should be flagged as missing evidence rather than as a circular step. Similarly, the inconsistencies between Algorithms 1 and 2 and between Eqs. (10) and (11) are internal errors, not definitional circularity. Because the central claim is deferred rather than derived from its own outputs, the circularity burden is low; the correct finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The architecture introduces a handful of hand-chosen hyperparameters (bins, slots, loss weights, filter radii) and relies on unverified domain assumptions about the usefulness of frozen cue extractors and the trainability of uncertainty heads. The key-value memory bank is an invented component with no independent validation.

free parameters (5)
  • Number of depth bins K = 64
    Chosen by hand (Section 4.5); no sensitivity analysis provided.
  • Memory slots S and feature dim D = S=32, D=128
    Chosen by hand (Section 4.4); no ablation on memory size is reported except an unverified 3.1% ablation.
  • Loss weights alpha, beta, gamma = unspecified
    Weights in Eq. (5) are introduced but values are not stated in the paper.
  • Guided filter parameters r, epsilon = r=4, epsilon=1e-3
    Chosen by hand (Section 4.6).
  • Gated memory write initial eta = 0.1
    Initialized to 0.1 (Section 4.4).
assumptions (3)
  • domain assumption Pre-trained HED, SDPS-Net and HorizonNet provide useful, complementary monocular cues
    The whole pipeline rests on this; no experiment or analysis validates it in this version.
  • domain assumption The uncertainty heads on frozen networks can be trained without variance labels
    Section 4.1 appends variance heads but gives no supervision scheme for them.
  • ad hoc to paper The described memory read-write and fusion blocks are implementable as written
    Algorithm 1 initializes M to zeros and never updates it; Algorithm 2 and Eq. (10)-(11) give conflicting update formulas.
invented entities (1)
  • Key-value working memory bank (S=32, D=128)
    purpose: To store and reweight monocular cue combinations across transformer layers
    Novel module introduced by the paper; only evidence is an unshown 3.1% ablation in Section 8.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of THIRDEYE: Cue-Aware Monocular Depth Estimation via Brain-Inspired Multi-Stage Fusion." pith.science (2026). https://pith.science/paper/LM3LPFFO

@misc{pith2026250620877,
  author       = {Pith},
  title        = {Pith review of: THIRDEYE: Cue-Aware Monocular Depth Estimation via Brain-Inspired Multi-Stage Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LM3LPFFO}},
  note         = {Machine review of arXiv:2506.20877}
}
read the original abstract

Monocular depth estimation methods traditionally train deep models to infer depth directly from RGB pixels. This implicit learning often overlooks explicit monocular cues that the human visual system relies on, such as occlusion boundaries, shading, and perspective. Rather than expecting a network to discover these cues unaided, we present ThirdEye, a cue-aware pipeline that deliberately supplies each cue through specialised, pre-trained, and frozen networks. These cues are fused in a three-stage cortical hierarchy (V1->V2->V3) equipped with a key-value working-memory module that weights them by reliability. An adaptive-bins transformer head then produces a high-resolution disparity map. Because the cue experts are frozen, ThirdEye inherits large amounts of external supervision while requiring only modest fine-tuning. This extended version provides additional architectural detail, neuroscientific motivation, and an expanded experimental protocol; quantitative results will appear in a future revision.

Figures

Figures reproduced from arXiv: 2506.20877 by the authors.

Figure 1
Figure 1. High Level Overview of Pipeline 1. Cue Extraction (Frozen Specialists). The input RGB image I ∈R H×W×3 is first analysed by three light-weight networks—HED, SDPS-Net, and HorizonNet—delivering edge maps E, surface normals N, global layout P, and their associated uncertainties σE,N,P . All specialists are queried once and cached. 2. V1 Feature Stem. Raw RGB channels and down-sampled cue maps are concatenated and pass… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 48 canonical work pages

  1. [8]

    Nasr and A

    J. Nasr and A. Khosla. ‘Edge-based Cues Align Depth Estimation with Human Perception.” CVPR, 2021

    work page 2021

  2. [1]

    Eigen, C

    D. Eigen, C. Puhrsch, and R. Fergus. ‘Depth Map Prediction from a Single Image using a Multi -Scale Deep Network.” NIPS, 2014

    work page 2014

  3. [2]

    Schneegans et al

    E. Schneegans et al. ‘Feedback scales the spatial tuning of cortical responses during both perception and memory.” PNAS, 2024

    work page 2024

  4. [3]

    Linde-Domingo et al

    C. Linde-Domingo et al. ‘Cortical feedback loops bind distributed representations of working memory.”Nature, 2022

    work page 2022

  5. [4]

    J. X. Yu et al. ‘Working memory signals in early visual cortex mediate distraction.” Nature Communications, 2022

    work page 2022

  6. [5]

    Noudoost and T

    B. Noudoost and T. Moore. ‘Neuromodulation of Persistent Activity and Working Memory.”Frontiers in Neural Circuits, 2021

    work page 2021

  7. [6]

    Kendall and Y

    A. Kendall and Y . Gal. ‘What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?”NIPS, 2017

    work page 2017

  8. [7]

    Banerjee and P

    S. Banerjee and P. K. Matteucci. ‘Synergistic Integration of Monocular Cues for Depth Perception.” Vision Research, 2020

    work page 2020

Show all 67 references
  1. [9]

    Zamir et al

    I. Zamir et al. ‘Robust Monocular Depth Under Domain Shift via Specialist Fusion.” ICCV, 2023

    2023

  2. [10]

    Kim et al

    M. Kim et al. ‘FocusDepth: Leveraging Defocus Blur for Self-Supervised Monocular Depth.” CVPR, 2024

    2024

  3. [11]

    Depth map prediction from a single image using a multi-scale deep network,

    D. Eigen, C. Puhrsch, and R. Fergus, “Depth map prediction from a single image using a multi-scale deep network,” in *Advances in Neural Information Processing Systems* (NeurIPS), 2014

    2014

  4. [12]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” in *Proc. IEEE/CVF Int. Conf. on Computer Vision (ICCV)*, 2021, pp. 12179–12188

    2021

  5. [13]

    AdaBins: Depth estimation using adaptive bins,

    S. F. Bhat, I. Alhashim, and P. Wonka, “AdaBins: Depth estimation using adaptive bins,” in *Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)*, 2021, pp. 4009–4018

    2021

  6. [14]

    ZoeDepth: Zero-shot transfer by combining relative and metric depth,

    S. F. Bhat, R. Birkl, D. Wofk, P. Wonka, and M. Müller, “ZoeDepth: Zero-shot transfer by combining relative and metric depth,” *arXiv preprint* arXiv:2302.12288, 2023

    2023 arXiv

  7. [15]

    Holistically-nested edge detection,

    S. Xie and Z. Tu, “Holistically-nested edge detection,” in *Proc. IEEE Int. Conf. on Computer Vision (ICCV)*, 2015, pp. 1395–1403

    2015

  8. [16]

    Self-calibrating deep photometric stereo networks,

    G. Chen, K. Han, B. Shi, Y . Matsushita, and K.-Y . K. Wong, “Self-calibrating deep photometric stereo networks,” in *Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)*, 2019, pp. 8739–8748

    2019

  9. [17]

    HorizonNet: Learning room layout with 1D representation and pano stretch data augmentation,

    C. Sun, C.-W. Hsiao, M. Sun, and H.-T. Chen, “HorizonNet: Learning room layout with 1D representation and pano stretch data augmentation,” in *Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)*, 2019, pp. 1047–1056

    2019

  10. [18]

    Vision meets robotics: The KITTI dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,” *International Journal of Robotics Research*, vol. 32, no. 11, pp. 1231–1237, 2013, doi:10.1177/0278364913491297

    2013 doi

  11. [19]

    MegaDepth: Learning single-view depth prediction from internet photos,

    Z. Li and N. Snavely, “MegaDepth: Learning single-view depth prediction from internet photos,” in *Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)*, 2018, pp. 2041–2050

    2018

  12. [20]

    TartanAir: A dataset to push the limits of visual SLAM,

    W. Wang et al., “TartanAir: A dataset to push the limits of visual SLAM,” *arXiv preprint* arXiv:2003.14338, 2020

    2003 arXiv

  13. [21]

    Humans integrate visual and haptic information in a statistically optimal fashion,

    M. O. Ernst and M. S. Banks, “Humans integrate visual and haptic information in a statistically optimal fashion,” *Nature*, vol. 415, pp. 429–433, 2002, doi:10.1038/415429a

    2002 doi

  14. [22]

    Synaptic theory of working memory,

    G. Mongillo, O. Barak, and M. Tsodyks, “Synaptic theory of working memory,” *Science*, vol. 319, pp. 1543–1546, 2008, doi:10.1126/science.1150769

    2008 doi

  15. [24]

    Working memory representa- tions in visual cortex mediate distraction effects,

    G. E. Hallenbeck, T. C. Sprague, M. Rahmati, K. K. Sreenivasan, and C. E. Curtis, “Working memory representa- tions in visual cortex mediate distraction effects,” *Nature Communications*, vol. 12, 2021, doi:10.1038/s41467- 021-24973-1

    2021 doi

  16. [25]

    Cortical feedback loops bind distributed representations of working memory,

    I. V oitov and T. D. Mrsic-Flogel, “Cortical feedback loops bind distributed representations of working memory,” *Nature*, vol. 608, pp. 381–389, 2022, doi:10.1038/s41586-022-05014-3. 11 A PREPRINT - JULY 25, 2025

    2022 doi

  17. [26]

    Feedback scales the spatial tuning of cortical responses during both visual working memory and long-term memory,

    R. Woodry, J. Curtis, and D. Winawer, “Feedback scales the spatial tuning of cortical responses during both visual working memory and long-term memory,” *Journal of Neuroscience*, 2025, doi:10.1523/JNEUROSCI.0681- 24.2025

    2025 doi

  18. [27]

    Neuromodulation of persistent activity and working memory circuitry in primate prefrontal cortex by muscarinic receptors,

    S. Vijayraghavan and S. Everling, “Neuromodulation of persistent activity and working memory circuitry in primate prefrontal cortex by muscarinic receptors,” *Frontiers in Neural Circuits*, vol. 15, 2021, art. 648624, doi:10.3389/fncir.2021.648624

    2021

  19. [28]

    What uncertainties do we need in Bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?” in *Advances in Neural Information Processing Systems* (NeurIPS), 2017

    2017

  20. [29]

    On the uncertainty of self-supervised monocular depth estimation,

    M. Poggi, F. Aleotti, F. Tosi, and S. Mattoccia, “On the uncertainty of self-supervised monocular depth estimation,” in *Proc. IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR)*, 2020

    2020

  21. [30]

    Recurrent models of visual attention,

    V . Mnih, N. Heess, A. Graves, and K. Kavukcuoglu, “Recurrent models of visual attention,” in *Advances in Neural Information Processing Systems* (NeurIPS), 2014

    2014

  22. [32]

    Vision Transformers for Dense Prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision Transformers for Dense Prediction,” *arXiv preprint* arXiv:2103.13413, 2021

    2021 arXiv

  23. [33]

    Densely Connected Convolutional Networks,

    G. Huang, Z. Liu, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” *arXiv preprint* arXiv:1608.06993, 2016

    2016 arXiv

  24. [34]

    Depth from a Single Image by Harmonizing Overcomplete Local Network Predictions,

    A. Chakrabarti, J. Shao, and G. Shakhnarovich, “Depth from a Single Image by Harmonizing Overcomplete Local Network Predictions,” *arXiv preprint* arXiv:1605.07081, 2016

    2016 arXiv

  25. [35]

    Look deeper into depth: Monocular depth estimation with semantic booster and attention-driven loss,

    J. Jiao, Y . Cao, Y . Song, and R. Lau, “Look deeper into depth: Monocular depth estimation with semantic booster and attention-driven loss,” in *Proc. European Conference on Computer Vision (ECCV)*, pp. 55–71, 2018

    2018

  26. [36]

    AdaBins: Depth Estimation Using Adaptive Bins,

    S. F. Bhat, I. Alhashim, and P. Wonka, “AdaBins: Depth Estimation Using Adaptive Bins,” in *Proc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, pp. 4008–4017, 2021

    2021

  27. [37]

    Uncertainty-guided annotation enhances segmentation with the human-in-the-loop,

    N. Khalili, J. Spronck, F. Ciompi, J. van der Laak, and G. Litjens, “Uncertainty-guided annotation enhances segmentation with the human-in-the-loop,” *arXiv preprint* arXiv:2404.07208, 2024

    2024 arXiv

  28. [38]

    Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs,

    J. Kubilius, M. Schrimpf, H. Hong, N. J. Majaj, R. Rajalingham, E. B. Issa, K. Kar, P. Bashivan, J. Prescott-Roy, K. Schmidt, A. Nayebi, D. Bear, D. L. K. Yamins, and J. J. DiCarlo, “Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs,” *arXiv preprint* a...

    1909 arXiv

  29. [39]

    Recurrent Models of Visual Attention,

    V . Mnih, N. Heess, A. Graves, and K. Kavukcuoglu, “Recurrent Models of Visual Attention,” *arXiv preprint* arXiv:1406.6247, 2014

    2014 arXiv

  30. [40]

    Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning,

    W. Lotter, G. Kreiman, and D. Cox, “Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning,” *arXiv preprint* arXiv:1605.08104, 2017

    2017 arXiv

  31. [41]

    Neural Turing Machines,

    A. Graves, G. Wayne, and I. Danihelka, “Neural Turing Machines,” *arXiv preprint* arXiv:1410.5401, 2014

    2014 arXiv

  32. [42]

    One-shot Learning with Memory-Augmented Neural Networks,

    A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap, “One-shot Learning with Memory-Augmented Neural Networks,” *arXiv preprint* arXiv:1605.06065, 2016

    2016 arXiv

  33. [43]

    Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition,

    Q. Hou, Z. Jiang, L. Yuan, M.-M. Cheng, S. Yan, and J. Feng, “Vision Permutator: A Permutable MLP-Like Architecture for Visual Recognition,” *arXiv preprint* arXiv:2106.12368, 2021

    2021 arXiv

  34. [44]

    Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks,” in *Proc. IEEE International Conference on Computer Vision (ICCV)*, 2017

    2017

  35. [45]

    One-shot Entropy Minimization,

    Z. Gao, L. Chen, J. Zhou, and B. Dai, “One-shot Entropy Minimization,” *arXiv preprint* arXiv:2505.20282, 2025

    2025 arXiv

  36. [46]

    D. H. Hubel and T. N. Wiesel, ‘Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex,” Journal of Physiology, vol. 160, pp. 106–154, 1962

    1962

  37. [47]

    C. Han, W. Huang, Y . R. Su, Z. J. He, and T. L. Ooi, ‘Evidence in support of the border-ownership neurons for representing textured figures,” iScience, vol. 23, no. 8, art. 101394, 2020, doi:10.1016/j.isci.2020.101394

    2020

  38. [48]

    D. Y . Tsao, W. Vanduffel, R. B. H. Tootell, P. Fize, and G. A. Orban, ‘Tuning of macaque V2 neurons for binocular disparity and border ownership,” Science, vol. 299, no. 5606, pp. 417–420, 2003

    2003

  39. [49]

    M. R. Warden and E. K. Miller, ‘Task-dependent changes in short-term memory in the prefrontal cortex,”Journal of Neuroscience, vol. 30, no. 47, pp. 15801–15810, 2010. 12 A PREPRINT - JULY 25, 2025

    2010

  40. [50]

    G. C. DeAngelis and W. T. Newsome, ‘Organization of disparity-selective neurons in macaque area MT,”Journal of Neuroscience, vol. 19, no. 4, pp. 1398–1415, 1999, doi:10.1523/JNEUROSCI.19-04-01398.1999

    1999

  41. [51]

    M. S. Livingstone and D. H. Hubel, ‘Anatomy and physiology of a color system in the primate visual cortex,” Journal of Neuroscience, vol. 4, no. 1, pp. 309–356, 1984

    1984

  42. [52]

    R. J. van Beers, A. C. Sittig, and J. J. van der Gon, ‘Integration of proprioceptive and visual position informa- tion: An experimentally supported model,” Journal of Neurophysiology, vol. 81, no. 3, pp. 1355–1364, 2002, doi:10.1152/jn.1999.81.3.1355

    2002 doi

  43. [53]

    Neural correlations, population coding and computation,

    B. B. Averbeck, P. E. Latham, and A. Pouget, “Neural correlations, population coding and computation,” *Nat. Rev. Neurosci.*, vol. 7, no. 5, pp. 358–366, 2006

    2006

  44. [54]

    I. Fine, B. Finnegan, and D. J. Fine, ‘The neural plasticity and functional organization of amblyopia: Review of current knowledge,” Documenta Ophthalmologica, vol. 113, no. 2, pp. 115–128, 2006, doi:10.1007/s10633-006- 9014-6

    2006 doi

  45. [55]

    D. H. Baker, ‘Amblyopia and binocular vision,” Current Biology, vol. 30, no. 24, pp. R1372–R1374, 2020, doi:10.1016/j.cub.2020.10.030

    2020 doi

  46. [56]

    X. Wang, Y . Li, and S. Zhang, ‘Reorganization of visual cortical circuits for motion-parallax depth perception in monocularly reared mice,” Nature Neuroscience, vol. 23, pp. 1144–1151, 2020, doi:10.1038/s41593-020-0669-9

    2020 doi

  47. [57]

    Bridge, P

    H. Bridge, P. Cicmil, S. Cowie, and A. Parker, ‘Visual cortical plasticity and recovery after early visual deprivation,” NeuroImage, vol. 78, pp. 353–360, 2013, doi:10.1016/j.neuroimage.2013.03.016

    2013 doi

  48. [58]

    Nassi, J

    J. Nassi, J. N. Gomez, and M. S. Livingstone, ‘Blur gradient coding by macaque MT neurons adapts with binocular correlation,” Journal of Neuroscience, vol. 41, no. 29, pp. 6129–6141, 2021, doi:10.1523/JNEUROSCI.0361- 21.2021

    2021 doi

  49. [59]

    P. S. Goldman -Rakic, ‘Cellular basis of working memory,” emphNeuron, vol. 14, no. 3, pp. 477–485, 1995, doi:10.1016/0896-6273(95)90304-6

    1995 doi

  50. [60]

    E. K. Miller and T. J. Buschman, ‘Working memory-like activity in monkey visual cortex,” emphScience, vol. 361, no. 6406, pp. 83–87, 2018, doi:10.1126/science.aao4704

    2018 doi

  51. [61]

    S. A. Harrison and F. Tong, ‘Decoding reveals the contents of visual working memory in early visual areas,” emphNature, vol. 458, pp. 632–635, 2009, doi:10.1038/nature07832

    2009 doi

  52. [62]

    M. G. Stokes, ‘’Activity-silent’ working memory in prefrontal cortex: A dynamic coding framework,” emphTrends in Cognitive Sciences, vol. 19, no. 7, pp. 394–405, 2015, doi:10.1016/j.tics.2015.05.004

    2015 doi

  53. [63]

    M. Mank, A. Fischer, T. Köster, J. E. Westkott, and A. Griesbeck, ‘High-resolution calcium imaging of neuronal activity in behaving mice,” Neuron, vol. 60, no. 6, pp. 1068–1079, 2008, doi:10.1016/j.neuron.2008.10.028

    2008 doi

  54. [64]

    K. J. Alcock, L. B. Smith, and M. G. Stokes, ‘Silent substrates of visual working memory revealed by layer-specific fMRI,” NeuroImage, vol. 282, art. 119276, 2024, doi:10.1016/j.neuroimage.2023.119276

    2024

  55. [65]

    M. E. Hasselmo and B. P. McGaughy, ‘High acetylcholine levels set circuit dynamics for attention and encoding and low acetylcholine levels set dynamics for consolidation,” Progress in Brain Research, vol. 145, pp. 207–231, 2004, doi:10.1016/S0079-6123(03)45015-2

    2004 doi

  56. [66]

    Wang, ‘The prefrontal cortex as a quintessential cognitive-type’ neural circuit: Working memory and decision making,” in Principles of Frontal Lobe Function, 2nd ed., pp

    X.-J. Wang, ‘The prefrontal cortex as a quintessential cognitive-type’ neural circuit: Working memory and decision making,” in Principles of Frontal Lobe Function, 2nd ed., pp. 226–248, Oxford University Press, 2013

    2013

  57. [67]

    Kirkpatrick et al

    J. Kirkpatrick et al. , ‘Overcoming catastrophic forgetting in neural networks,” Proceedings of the National Academy of Sciences, vol. 114, no. 13, pp. 3521–3526, 2017, doi:10.1073/pnas.1611835114

    2017 doi

  58. [68]

    Y . Zou, Z. Lv, X. Wang, R. Yu, S. Zhang, and A. Gaidon, ‘DDAD: The Diverse Driving Dataset,”arXiv preprint arXiv:2002.10303, 2020

    2002 arXiv

  59. [69]

    Silberman, D

    N. Silberman, D. Hoiem, P. Kohli, and R. Fergus, ‘Indoor segmentation and support inference from RGB-D images,” in Proc. European Conference on Computer Vision (ECCV), 2012, pp. 746–760. 13

    2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.