Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

EvidenceMoE: A Physics-Guided Mixture-of-Experts with Evidential Critics for Advancing Fluorescence Light Detection and Ranging in Scattering Media

T0 review · 4 major / 7 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 EvidenceMoE estimates depth and fluorescence lifetime directly from raw FLiDAR signals in scattering media, reaching NRMSE of 0.030 and 0.074 on simulated data.

desk verdict A sensible FLiDAR depth/lifetime architecture with a real reliability-calibration gap and no baselines; deserves review but needs major empirical revision. read the letter →

arxiv 2505.21532 v1 pith:LQ5LCAQF submitted 2025-05-23 cs.CV cs.AIcs.LGphysics.optics

Ismail Erbas , Ferhat Demirkiran , Karthik Swaminathan , Naigang Wang , Navid Ibtehaj Nizam , Stefan T. Radev , Kaoutar El Maghraoui , Xavier Intes
show 1 more author
Vikas Pandey
This is my paper · ORCID
classification cs.CVcs.AIcs.LGphysics.optics
keywords fluorescencelifetimeimagingLiDARscatteringmediamixture-of-expertsevidentialdeeplearninguncertaintyquantificationdepthestimationMonteCarlosimulation
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

The paper aims to show that depth and fluorescence lifetime can be recovered directly from raw, time-resolved fluorescence LiDAR signals in scattering media, where photon scattering blurs the boundary between time-of-flight and fluorescence decay. The proposed EvidenceMoE architecture assigns separate expert networks to the early, late, and full temporal segments of the signal, matching the physics that early photon arrivals mostly encode depth and the later decay mostly encodes lifetime. Each expert is paired with an evidence-based critic that outputs a reliability score and a correction signal, and a Decider network uses these scores to fuse the corrected expert predictions into a final estimate. On Monte-Carlo-simulated FLiDAR data for tumor depth detection in scattering tissue, the framework reports a normalized root mean squared error of 0.030 for depth and 0.074 for fluorescence lifetime, with critic quality scores near 95% and 96.5% that reflect the relative precision of the two estimates. A sympathetic reader would care because direct end-to-end estimation could remove the need for separate optical-property measurements and intermediate analytical inversions in clinical FLiDAR.

What carries the argument

The load-bearing mechanism is the pairing of physics-guided temporal experts with Evidence-Based Dirichlet Critics (EDCs). Early, Late, and Global experts process different temporal slices of the signal, reflecting the physics that early photons encode target depth while late decay mostly encodes fluorescence lifetime; each EDC takes the expert's pooled features and auxiliary prediction and outputs the parameters of a product of independent Beta distributions, whose mean acts as a quality score, plus a correction residual. A Decider network then gates the corrected expert predictions with weights derived from these quality scores and the global expert's pooled features. This arrangement turns uncertainty about an expert's reliability into an explicit, differentiable signal that shapes the final estimate.

What would settle it

Run the same trained architecture on time-resolved fluorescence measurements from a tissue-mimicking phantom with independently known fluorophore depth and lifetime; if the depth or lifetime error is substantially worse than the reported NRMSE of 0.030 and 0.074, or the EDC quality scores stop tracking actual error, the claim that simulation-trained results transfer to real scattering media is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a physics-guided mixture-of-experts network, with each expert constrained to a physically meaningful temporal segment of the FLiDAR signal and each expert vetted by an evidential critic, can disentangle the convolved effects of scattering, time-of-flight, and fluorescence decay without separately estimating medium optical properties. Each expert produces an auxiliary prediction and pooled attention features; its paired Evidence-Based Dirichlet Critic predicts the parameters of independent Beta distributions, whose mean is a per-expert quality score, and also outputs a residual correction. A Decider network gates the corrected expert predictions using these quality scores and the global expert's features to produce the final depth and lifetime estimates. In realistic simulations of fluorescence LiDAR in scattering tissue, the full model reaches NRMSE of 0.030 for depth and 0.074 for fluorescence lifetime, with mean critic quality scores near 95% for depth and 96.5% for lifetime that align with the observed precision ordering.

Load-bearing premise

The results rest on the assumption that the computer-simulated FLiDAR signals, including tissue scattering, detector response, and noise, faithfully match real fluorescence LiDAR signals in scattering media; if that match fails, the reported depth and lifetime errors and the quality scores may not transfer to real measurements.

Editorial extensions

If this is right

  • Depth and fluorescence lifetime could be recovered from a single raw temporal FLiDAR acquisition, without separate measurement of tissue optical properties, supporting real-time clinical use.
  • The EDC quality scores provide a per-input reliability indicator that could flag low-confidence depth or lifetime estimates during surgical guidance.
  • The early/late/global temporal split could transfer to other time-resolved sensing domains where different physical parameters dominate different parts of the recorded signal.
  • The reported NRMSE values of 0.030 for depth and 0.074 for lifetime, together with the ablation results, set a benchmark on simulated scattering-media FLiDAR data against which subsequent models can be compared.

Reading between the lines

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

  • Editorial inference: if the EDC quality scores are calibrated on held-out noise or scattering levels, they could serve as a learned confidence map that tells a surgeon when a depth estimate is trustworthy.
  • Editorial inference: the early/late/global segmentation is not specific to fluorescence; it could be applied to acoustic, radar, or spectral time-series where early arrivals carry geometry and late arrivals carry material properties.
  • Editorial inference: the natural decisive test is a tissue-phantom or ex vivo experiment with independently known depth and lifetime; if the reported accuracy and quality-score calibration survive the simulation-to-real shift, the clinical case is more convincing.
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 / 7 minor

Summary. The paper proposes EvidenceMoE, an end-to-end deep learning framework for estimating depth and fluorescence lifetime directly from raw time-resolved FLiDAR signals in scattering media. The architecture consists of three physics-guided experts (early, late, and global temporal segments), each paired with an Evidence-Based Dirichlet Critic (EDC) that outputs quality scores and correction signals, followed by a Decider network that performs quality-weighted fusion. The method is trained and evaluated on MCX-simulated data for non-invasive cancer-cell depth detection, reporting a depth NRMSE of 0.030 and a lifetime NRMSE of 0.074. The paper includes ablation studies in Tables 1 and 2 and visualizations of attention maps to support the claim of expert specialization.

Significance. If the claims are substantiated, the paper would make a useful contribution to FLiDAR parameter estimation in scattering media: it introduces a concrete physics-guided temporal inductive bias, a lightweight evidence-based critic mechanism, and an adaptive fusion scheme for a clinically relevant ill-posed inverse problem. The MCX-based simulation pipeline and the ablation suite are constructive elements. However, the current experimental evidence does not yet establish the central claims: the reliability scores are not shown to calibrate individual predictions, the headline NRMSE values are not defined, the full model does not consistently beat its own ablations on the lifetime target, and there is no external baseline or experimental validation. The significance is therefore conditional on addressing these issues.

major comments (4)
  1. [§3.3, Eq. (1); §8, Eqs. (25)–(29)] The quality score is supervised only through q_target = 1/(1+κ·MAE), where MAE is computed over the entire batch (Eq. 25). Lquality (Eq. 29) therefore trains the per-sample critic outputs q_{k,d} to match a single batch-level scalar. With a batch size of 512, an individual sample contributes about 0.2% of the target, so the critic is never told whether that sample is more or less reliable than its batchmates. The high average Q.Depth and Q.Life scores in Table 1 and Figure 4c–d consequently only demonstrate that the model can reproduce mean expert error, not that it calibrates individual predictions. Since per-prediction reliability is a stated contribution and a clinical motivation (Sec. 1), this is a load-bearing gap. Please provide a per-sample reliability target (e.g., per-sample absolute error) or a calibration/selective-prediction evaluation on held-out samples.
  2. [§4.2; Abstract] The headline metrics D.NRMSE and L.NRMSE(f) are never defined. The denominator of the normalization is not stated, so the numbers 0.030 and 0.074 cannot be interpreted or compared with future work. For example, if the normalization is the standard deviation of the ground-truth values, the values are dimensionless; if it is the range, the values are different. Please specify the exact normalization for both metrics and, if possible, report unnormalized RMSE as well.
  3. [Table 1; §4.5.1] The full model does not outperform its own ablations on the lifetime target: L.NRMSE(f) is 0.074±0.022 for the full model, 0.063±0.018 for 'No evidential correction', and 0.063±0.051 for 'Heteroscedastic experts only'. On this single test set, the evidential correction appears to degrade lifetime estimation. No significance tests, seed counts, or error bars across independent runs are provided, so statements in §4.5.1 that EDCs make a 'substantial contribution' to overall efficacy are not supported by the reported numbers. Please either show consistent improvement over multiple seeds and with proper statistical testing, or temper the claim.
  4. [§4.1, §4.5] All results are obtained on MCX-simulated data with no validation on experimentally measured FLiDAR signals. The paper also does not compare against an external baseline such as a standard deep regression model, an analytical inversion method, or established uncertainty quantification baselines (MC dropout, deep ensembles). The claim of 'advancing' FLiDAR estimation therefore rests on the fidelity of the simulation and on relative improvements that are not demonstrated. At minimum, include a non-MoE baseline trained on the same data and a discussion of simulation-to-real domain shift.
minor comments (7)
  1. [§5] The Conclusion says 'NRMSE over 0.03 and 0.074' (Sec. 5); since lower NRMSE is better, this should read 'NRMSE of 0.030 and 0.074'.
  2. [Throughout] The paper alternates between 'EvidenceMoe' and 'EvidenceMoE' (e.g., title vs. Sec. 5); standardize the capitalization.
  3. [§7.2] Section 7.2 calls the critic a 'Dirichlet' critic, but the model outputs a product of independent Beta distributions per dimension, not a single Dirichlet; clarify the terminology.
  4. [§8, Eq. (26)] Equation (26) uses ϵ in q_gt while Eq. (1) does not; define ϵ consistently in both places.
  5. [Figure 4] The caption of Figure 4 is minimal; add axis labels, units (cm, ns), and a description of what the plotted points and error bars (if any) represent.
  6. [Table 2] Table 2 includes a 'Uniform gating' row and a 'No auxiliary MAE' row that are not discussed in the main text; add a sentence explaining these variants.
  7. [Reproducibility] No code or data availability statement is included; for reproducibility, please add one.

Circularity Check

1 steps flagged · score 6.0 of 10

EDC quality scores are trained to a self-defined batch-MAE target (Eq. 1, Eqs. 25-29), so the reported 95-96.5% reliability scores reduce to the critic fitting its own target rather than independent per-sample calibration.

  1. fitted input called prediction [Section 3.3, Eq. (1); Section 8.1, Eqs. (25)-(29); Section 4.5]
    "We compute the target quality score as the inverse of the expert's Mean Absolute Error (MAE) on each output dimension. For each dimension d, the target score is given by qtarget_{k,d} = 1/(1+κ·MAE_{k,d}) with MAE_{k,d} = 1/N Σ_{i=1}^N |y^{(i)}_{aux,k,d} − y^{(i)}_{true,k′,d}|, (1) ... This formulation enables the critic to learn a continuous notion of reliability directly grounded in the expert's observed performance."

    The EDC quality score is trained, via Eq. (29), to regress toward qgt,k,d = (1+κ·MAE_{k,d}+ε)^(−1), where MAE_{k,d} is averaged over the entire batch of the expert's own auxiliary predictions (Eqs. 25-26). Thus q is, by construction, a learned approximation of a deterministic function of the expert's batch error; it is not an independent per-sample calibration. Consequently, the Section 4.5 statement that average Q.Depth≈95% and Q.Life≈96.5% 'underscores the utility of the EDC-generated quality scores in reflecting prediction reliability' only demonstrates that the critic tracks its own self-defined target. Because a single batch-level scalar target is shared by all N samples, the critic is never supervised on whether one sample is more reliable than another.

full rationale

The central depth/lifetime regression is not circular: the experts and final decider are trained with L1 losses against MCX-simulated ground-truth depth and lifetime (Eqs. 23-24), and the reported NRMSEs are computed on held-out test samples from the same simulation distribution. The physics-guided early/late specialization is an inductive bias, not a fitted target. However, one load-bearing claimed contribution—the EDC quality scores as reliability indicators—is partially circular. The critic target is defined from the expert's own batch MAE (Eq. 1; Eqs. 25-26), so the reported ~95-96% quality scores are the model reproducing a self-defined function of its training error. This does not corrupt the depth/lifetime NRMSE, but it does mean the paper's confidence-calibration claim is not independently evidenced. Self-citations to the authors' MC data generator ([23,24,26]) are used for simulation workflows but are not invoked to forbid alternatives or to justify the main loss; they do not add circularity beyond the quality-score target.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests primarily on the realism of the MCX simulation, the early and late temporal decomposition, and a self-defined quality-target loss. No new physical entities are introduced. The EDC and Decider are architectural components whose validation is internal to the paper.

free parameters (6)
  • kappa in quality target q_target = 1/(1+kappa*MAE) = 2 (8 in ablation)
    Chosen by the authors to control how error maps to target quality; it changes critic training and affects final errors, with the kappa=8 row giving D.NRMSE 0.040 versus 0.030.
  • lambda_damp damping factor = d = 0.5 and 1.0 tested
    Scales the correction signal in L_corr (Eq. 30); ablation rows show different NRMSE, so the result depends on this choice.
  • loss weights lambda_primary, lambda_aux, lambda_quality, lambda_corr, lambda_pen, gamma, lambda_KL = not reported
    Appear in L_total (Eq. 22) and in the evidence-weighted penalty (Eq. 32); their values are not given, so the training objective is under-specified.
  • temporal segment boundary (early vs late split) = first half / second half, exact bins unspecified
    This is the basis of the physics-guided specialization; no sensitivity analysis with respect to the split point is provided.
  • phased training epochs N1, N2, N3 = varied across ablations (e.g., 1/6, 5/15, 3/8, 10/10)
    The phased training schedule is manually tuned and affects reported accuracy.
  • expert encoder hyperparameters (hidden size H, kernel size K, transformer depth, MLP depth) = not fully specified in paper
    The architecture description in Section 7 leaves these values unspecified; they determine model capacity and are hand-chosen.
assumptions (5)
  • domain assumption MCX simulation of the radiative transport equation with IRF and noise yields FLiDAR signals realistic enough for clinical validation.
    Section 4.1 states data are generated with MCX using camera IRF and noise; no experimental validation is provided.
  • domain assumption Early photons primarily encode depth and late photons primarily encode fluorescence lifetime.
    Section 6 and ref. [36]; this is the basis for assigning expert roles.
  • ad hoc to paper q_target = 1/(1+kappa*MAE) computed from the expert's own batch error is a valid reliability proxy.
    Eq. 1 and Eq. 26; the reliability claims inherit this self-defined target.
  • standard math Beta distributions with softplus evidence are a well-defined evidential model for regression quality.
    Section 7.2; standard EDL machinery, but calibration is not tested.
  • domain assumption The training loss and phased schedule lead to generalization from simulated to real signals.
    Sections 5 and 8; no out-of-distribution or real-data evaluation is included.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EvidenceMoE: A Physics-Guided Mixture-of-Experts with Evidential Critics for Advancing Fluorescence Light Detection and Ranging in Scattering Media." pith.science (2026). https://pith.science/paper/LQ5LCAQF

@misc{pith2026250521532,
  author       = {Pith},
  title        = {Pith review of: EvidenceMoE: A Physics-Guided Mixture-of-Experts with Evidential Critics for Advancing Fluorescence Light Detection and Ranging in Scattering Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQ5LCAQF}},
  note         = {Machine review of arXiv:2505.21532}
}
read the original abstract

Fluorescence LiDAR (FLiDAR), a Light Detection and Ranging (LiDAR) technology employed for distance and depth estimation across medical, automotive, and other fields, encounters significant computational challenges in scattering media. The complex nature of the acquired FLiDAR signal, particularly in such environments, makes isolating photon time-of-flight (related to target depth) and intrinsic fluorescence lifetime exceptionally difficult, thus limiting the effectiveness of current analytical and computational methodologies. To overcome this limitation, we present a Physics-Guided Mixture-of-Experts (MoE) framework tailored for specialized modeling of diverse temporal components. In contrast to the conventional MoE approaches our expert models are informed by underlying physics, such as the radiative transport equation governing photon propagation in scattering media. Central to our approach is EvidenceMoE, which integrates Evidence-Based Dirichlet Critics (EDCs). These critic models assess the reliability of each expert's output by providing per-expert quality scores and corrective feedback. A Decider Network then leverages this information to fuse expert predictions into a robust final estimate adaptively. We validate our method using realistically simulated Fluorescence LiDAR (FLiDAR) data for non-invasive cancer cell depth detection generated from photon transport models in tissue. Our framework demonstrates strong performance, achieving a normalized root mean squared error (NRMSE) of 0.030 for depth estimation and 0.074 for fluorescence lifetime.

Figures

Figures reproduced from arXiv: 2505.21532 by the authors.

Figure 1
Figure 1. End-to-end EvidenceMoE Workflow for FLiDAR-based Tumor Lifetime and Depth Estimation. FLiDAR signal acquisition begins with laser interaction with a fluorescent tumor target. The resulting emitted/scattered photons are captured by a time-resolved camera, yielding a complex temporal signal. EvidenceMoE architecture processes the captured signal to ultimately yield estimations for tumor depth and fluorescence lifetime… view at source ↗
Figure 2
Figure 2. Overall architecture of the Physics-guided EvidenceMoE model with Evidence Critics. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of Pooled Attention Weights Over Time for Both Depth and Lifetime Experts, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance results of EvidenceMoE model [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sticky Routing: Training MoE Models for Memory-Efficient Inference

    cs.LG 2026-06 conditional novelty 6.0 of 10

    A differentiable L2 consistency loss on consecutive MoE gate distributions reduces expert switch rate by up to 59% while preserving or improving perplexity, outperforming post-hoc router fine-tuning.

Reference graph

Works this paper leans on

45 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020

    Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression.Advances in neural information processing systems, 33:14927–14937, 2020

    2020

  2. [2]

    Weight uncertainty in neural network

    Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural network. InInternational conference on machine learning, pages 1613–1622. PMLR, 2015

    2015

  3. [3]

    Rensselaer Polytechnic Institute, 2012

    Jin Chen.Optical tomography in small animals with time-resolved Monte Carlo methods. Rensselaer Polytechnic Institute, 2012

    work page 2012

  4. [4]

    On the representation collapse of sparse mixture of experts.Advances in Neural Information Processing Systems, 35:34600–34613, 2022

    Zewen Chi, Li Dong, Shaohan Huang, Damai Dai, Shuming Ma, Barun Patra, Saksham Singhal, Payal Bajaj, Xia Song, Xian-Ling Mao, et al. On the representation collapse of sparse mixture of experts.Advances in Neural Information Processing Systems, 35:34600–34613, 2022

    work page 2022

  5. [5]

    Luminescence lifetime imaging of three-dimensional biological objects.Journal of Cell Science, 134(9):1–17, 2021

    Ruslan I Dmitriev, Xavier Intes, and Margarida M Barroso. Luminescence lifetime imaging of three-dimensional biological objects.Journal of Cell Science, 134(9):1–17, 2021

    work page 2021

  6. [6]

    Monte carlo simulation of photon migration in 3d turbid media accelerated by graphics processing units.Optics express, 17(22):20178–20190, 2009

    Qianqian Fang and David A Boas. Monte carlo simulation of photon migration in 3d turbid media accelerated by graphics processing units.Optics express, 17(22):20178–20190, 2009

    work page 2009

  7. [7]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016. 10

    2016

  8. [8]

    Inductive biases for deep learning of higher-level cognition

    Anirudh Goyal and Yoshua Bengio. Inductive biases for deep learning of higher-level cognition. Proceedings of the Royal Society A, 478(2266):20210068, 2022

    work page 2022

Show all 45 references
  1. [9]

    Ivo Grondman, Lucian Busoniu, Gabriel AD Lopes, and Robert Babuska. A survey of actor- critic reinforcement learning: Standard and natural policy gradients.IEEE Transactions on Systems, Man, and Cybernetics, part C (applications and reviews), 42(6):1291–1307, 2012

    2012

  2. [10]

    Analytical method for the fast time-domain reconstruction of fluorescent inclusions in vitro and in vivo.Biophysical journal, 98(2):350–357, 2010

    Sung-Ho Han, Salman Farshchi-Heydari, and David J Hall. Analytical method for the fast time-domain reconstruction of fluorescent inclusions in vitro and in vivo.Biophysical journal, 98(2):350–357, 2010

    2010

  3. [11]

    Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.Machine learning, 110(3):457–506, 2021

    Eyke Hüllermeier and Willem Waegeman. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods.Machine learning, 110(3):457–506, 2021

    2021

  4. [12]

    Hands-on bayesian neural networks—a tutorial for deep learning users.IEEE Computational Intelligence Magazine, 17(2):29–48, 2022

    Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, Wray Buntine, and Mohammed Ben- namoun. Hands-on bayesian neural networks—a tutorial for deep learning users.IEEE Computational Intelligence Magazine, 17(2):29–48, 2022

    2022

  5. [13]

    Application of lidar sensors for crop and working environment recognition in agriculture: A review.Remote Sensing, 16(24):4623, 2024

    Md Rejaul Karim, Md Nasim Reza, Hongbin Jin, Md Asrakul Haque, Kyu-Ho Lee, Joonjea Sung, and Sun-Ok Chung. Application of lidar sensors for crop and working environment recognition in agriculture: A review.Remote Sensing, 16(24):4623, 2024

    2024

  6. [14]

    Me-odal: Mixture- of-experts ensemble of cnn models for 3d object detection from automotive lidar point clouds

    Dhvani Katkoria, Jaya Sreevalsan-Nair, Mayank Sati, and Sunil Karunakaran. Me-odal: Mixture- of-experts ensemble of cnn models for 3d object detection from automotive lidar point clouds. In International Conference on Deep Learning Theory and Applications, pages 279–300. Sprin...

    2024

  7. [15]

    What uncertainties do we need in bayesian deep learning for computer vision?Advances in neural information processing systems, 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision?Advances in neural information processing systems, 30, 2017

    2017

  8. [16]

    Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles.Advances in neural information processing systems, 30, 2017

    2017

  9. [17]

    A review of tof-based lidar.Journal of Semiconductors, 45(10):101201, 2024

    Jie Ma, Shenglong Zhuo, Lei Qiu, Yuzhu Gao, Yifan Wu, Ming Zhong, Rui Bai, Miao Sun, and Patrick Yin Chiang. A review of tof-based lidar.Journal of Semiconductors, 45(10):101201, 2024

    2024

  10. [18]

    Mixture of experts: a literature survey.Artificial Intelligence Review, 42:275–293, 2014

    Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey.Artificial Intelligence Review, 42:275–293, 2014

    2014

  11. [19]

    Wide-field fluorescence lifetime imaging of cancer.Biomedical optics express, 1(2):627–640, 2010

    James McGinty, Neil P Galletly, Chris Dunsby, Ian Munro, Daniel S Elson, Jose Requejo-Isidro, Patrizia Cohen, Raida Ahmad, Amanda Forsyth, Andrew V Thillainayagam, et al. Wide-field fluorescence lifetime imaging of cancer.Biomedical optics express, 1(2):627–640, 2010

    2010

  12. [20]

    Noninvasive depth estimation using tissue optical properties and a dual-wavelength fluorescent molecular probe in vivo.Biomedical optics express, 8(6):3095–3109, 2017

    Jessica P Miller, Dolonchampa Maji, Jesse Lam, Bruce J Tromberg, and Samuel Achilefu. Noninvasive depth estimation using tissue optical properties and a dual-wavelength fluorescent molecular probe in vivo.Biomedical optics express, 8(6):3095–3109, 2017

    2017

  13. [21]

    The need for biases in learning generalizations

    Tom M Mitchell. The need for biases in learning generalizations. 1980

    1980

  14. [22]

    A comprehensive survey of mixture-of-experts: Algorithms, theory, and applications.arXiv preprint arXiv:2503.07137, 2025

    Siyuan Mu and Sen Lin. A comprehensive survey of mixture-of-experts: Algorithms, theory, and applications.arXiv preprint arXiv:2503.07137, 2025

    2025

  15. [23]

    Monte-carlo based data generator for fluorescence lifetime applications

    Navid Ibtehaj Nizam, Ismail Erbas, Vikas Pandey, and Xavier Intes. Monte-carlo based data generator for fluorescence lifetime applications. InOptical Tomography and Spectroscopy, pages JS4A–27. Optica Publishing Group, 2024

    2024

  16. [24]

    A novel technique for fluorescence lifetime tomography.bioRxiv, 2024

    Navid Ibtehaj Nizam, Vikas Pandey, Ismail Erbas, Jason T Smith, and Xavier Intes. A novel technique for fluorescence lifetime tomography.bioRxiv, 2024

    2024

  17. [25]

    Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processing systems, 32, 2019

    Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift.Advances in neural information processin...

    2019

  18. [26]

    Deep learning-based temporal deconvolution<? pag\break?> for photon time-of-flight distribution retrieval.Optics letters, 49(22):6457–6460, 2024

    Vikas Pandey, Ismail Erbas, Xavier Michalet, Arin Ulku, Claudio Bruschini, Edoardo Char- bon, Margarida Barroso, and Xavier Intes. Deep learning-based temporal deconvolution<? pag\break?> for photon time-of-flight distribution retrieval.Optics letters, 49(22):6457–6460, 2024

    2024

  19. [27]

    Arthur F Petusseau, Samuel S Streeter, Arin Ulku, Yichen Feng, Kimberley S Samkoe, Claudio Bruschini, Edoardo Charbon, Brian W Pogue, and Petr Bruza. Subsurface fluorescence time- of-flight imaging using a large-format single-photon avalanche diode sensor for tumor depth asses...

    2024

  20. [28]

    Evidential deep learning to quantify classification uncertainty.Advances in neural information processing systems, 31, 2018

    Murat Sensoy, Lance Kaplan, and Melih Kandemir. Evidential deep learning to quantify classification uncertainty.Advances in neural information processing systems, 31, 2018

    2018

  21. [29]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  22. [30]

    Time-moe: Billion-scale time series foundation models with mixture of experts.arXiv preprint arXiv:2409.16040, 2024

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts.arXiv preprint arXiv:2409.16040, 2024

    2024 arXiv

  23. [31]

    Exploring disordered light transport in scattering media to optimize random lasers.The Journal of Physical Chemistry C, 128(12):5321–5329, 2024

    Larissa A Silva, Fla´via S Ferreira, Gessineide S Oliveira, Andre L Moura, Rafael A de Oliveira, and Albert S Reyna. Exploring disordered light transport in scattering media to optimize random lasers.The Journal of Physical Chemistry C, 128(12):5321–5329, 2024

    2024

  24. [32]

    Macroscopic fluores- cence lifetime topography enhanced via spatial frequency domain imaging.Optics letters, 45 (15):4232–4235, 2020

    Jason T Smith, Enagnon Aguénounon, Sylvain Gioux, and Xavier Intes. Macroscopic fluores- cence lifetime topography enhanced via spatial frequency domain imaging.Optics letters, 45 (15):4232–4235, 2020

    2020

  25. [33]

    Fluorescence lifetime imaging for quantification of targeted drug delivery in varying tumor microenviron- ments.Advanced Science, 12(3):2403253, 2025

    Amit Verma, Vikas Pandey, Catherine Sherry, Taylor Humphrey, Christopher James, Kailie Mat- teson, Jason T Smith, Alena Rudkouskaya, Xavier Intes, and Margarida Barroso. Fluorescence lifetime imaging for quantification of targeted drug delivery in varying tumor microenviron- m...

    2025

  26. [34]

    Limoe: Mixture of lidar representation learners from automotive scenes.arXiv preprint arXiv:2501.04004, 2025

    Xiang Xu, Lingdong Kong, Hui Shuai, Liang Pan, Ziwei Liu, and Qingshan Liu. Limoe: Mixture of lidar representation learners from automotive scenes.arXiv preprint arXiv:2501.04004, 2025

    2025 arXiv

  27. [35]

    Antibody-target binding quantification in living tumors using macroscopy fluorescence lifetime forster resonance energy transfer imaging (mfli fret)

    Nanxue Yuan, Vikas Pandey, Amit Verma, John C Williams, Xavier Intes, and Margarida Barroso. Antibody-target binding quantification in living tumors using macroscopy fluorescence lifetime forster resonance energy transfer imaging (mfli fret). InVisualizing and Quantifying Drug...

    2024

  28. [36]

    L p regularization for early gate fluorescence molecular tomography.Optics letters, 39(14):4156–4159, 2014

    Lingling Zhao, He Yang, Wenxiang Cong, Ge Wang, and Xavier Intes. L p regularization for early gate fluorescence molecular tomography.Optics letters, 39(14):4156–4159, 2014. 12 6 Principles of Time-Resolved FLiDAR, Photon Scattering, and Fluorescence Lifetime Fluorescence LiDA...

    2014

  29. [37]

    Hybrid CNN-Transformer Encoder:The input segment xk ∈R Lk is initially processed by a sequence of 1D convolutional layers (kernel size K, with residual connections, Layer Normalization, and GELU activation). This allows the network to explicitly capture localized characteristi...

  30. [38]

    This layer learns attention weights over the time steps of henc k and computes a weighted sum, producing a single feature vector ϕk ∈R H

    Pooling Layer (fk):To obtain a fixed-size representation from the variable-length output sequence of the encoder, an Attention Pooling mechanism is employed. This layer learns attention weights over the time steps of henc k and computes a weighted sum, producing a single featu...

  31. [39]

    yaux,k =h k(ϕk;θ Ek )(6) Role of Global Expert Features:Beyond its auxiliary prediction yaux,g, the pooled feature vector ϕg from the global expert serves a dual purpose

    Auxiliary Prediction Head (hk):A small Multi-Layer Perceptron (MLP) with non-linear activation ( ReLU) maps the pooled feature vector ϕk to the expert’s specific auxiliary predictiony aux,k (dimensionD k = 1for early/late,D k = 2for global). yaux,k =h k(ϕk;θ Ek )(6) Role of Gl...

  32. [40]

    Shared Backbone (bk):The concatenated input zk is processed by a shared MLP backbone, bk. A relatively shallow architecture (two layers) with moderate hidden dimensions [32, 16] is employed to keep the critic computationally lightweight while providing sufficient capacity to e...

  33. [41]

    Evidence Head (evik):A dedicated linear layer, evik, maps the shared features hk to the raw evidence outputs ek ∈R 2×Dk (one positive and one negative evidence value per output dimensionD k of the corresponding expert). ek =evi k(hk;θ Ck )(9) To ensure the parameters of the re...

  34. [42]

    ∆k =corr k(hk;θ Ck )(12) The critic’s forward pass thus yields (αk, βk,∆ k)

    Correction Head (corrk):A separate linear layer maps hk to the correction signal ∆k ∈ RDk. ∆k =corr k(hk;θ Ck )(12) The critic’s forward pass thus yields (αk, βk,∆ k). The mean of the predicted Beta distribution serves as the point estimate for quality scoreq k: qk,d = αk,d αk...

  35. [43]

    Corrected Auxiliary Predictions (yaux):This is the set of auxiliary predictions from the experts, adjusted by the correction signals provided by the critics (Equation 12), yaux = {yaux,e ∈R 1, yaux,l ∈R 1, yaux,g ∈R 2}

  36. [44]

    Decider Feature (ϕg):This is the pooled feature vector ϕg ∈R H from the global expert’s encoder (Equation 5)

  37. [45]

    Full Quality Scores (qf ull):A vector containing the mean quality estimates derived from the Evidence Critics for all four quality dimensions: qf ull= [qe, ql, qg,d, qg,l]∈R 4 (Equa- tion 13). Gating Network (G):The core of the fusion mechanism is a learned gating network G, d...

Pith tools

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