arxiv: 2604.24236 · v1 · submitted 2026-04-27 · 📡 eess.IV · cs.AI· cs.CV· eess.SP

Recognition: unknown

Deep Learning-Enabled Dissolved Oxygen Sensing in Biofouling Environments for Ocean Monitoring

Adrien Desjardins, Manish K. Tiwari, Nikolaos Salaris

Authors on Pith no claims yet

Pith reviewed 2026-05-07 17:37 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CVeess.SP

keywords dissolved oxygen sensingbiofoulingphysics-informed neural networkvisual transformerocean monitoringStern-Volmer equationsensor driftdeep ensemble

0 comments

The pith

A visual transformer physics-informed neural network embeds the Stern-Volmer equation to sense dissolved oxygen with roughly 2 micromoles per liter error under biofouling.

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

The paper aims to show that inexpensive camera-based optoelectronic sensors can deliver reliable absolute dissolved oxygen readings in algae-fouled water by training a visual transformer model whose loss function directly incorporates the Stern-Volmer physical relation. This matters because widespread, low-cost DO monitoring is needed to track ocean ecosystem health and climate tipping points, yet biofouling has made such sensors impractical for long deployments. Training and testing on 14 days of accelerated biofouling in an algae-laden tank produces 92 percent and 89 percent lower mean average error than classical statistical or machine-learning baselines. A deep ensemble of the models also supplies uncertainty estimates that support self-diagnosis of sensor degradation.

Core claim

We introduce a sensing paradigm that combines camera-based DO sensors with a visual transformer (ViT)-based physics-informed neural network (PINN) for high-fidelity sensing under biofouling conditions. Training and testing data were obtained from an algae-laden water tank over 14 days to capture accelerated biofouling. The ViT-PINN, which embeds the Stern-Volmer (SV) equation into the loss function, reduces mean average error (MAE) by 92% and 89% compared to classical statistical and ML approaches, achieving ~2 umol/L absolute error. A deep ensemble further quantifies predictive uncertainty, enabling self-diagnostic sensing.

What carries the argument

The ViT-PINN, a visual transformer neural network whose loss function embeds the Stern-Volmer equation relating phosphorescence lifetime to oxygen concentration.

If this is right

Low-cost microstructured polymer-film sensors become viable for sustained ocean deployments without frequent cleaning.
Predictive uncertainty estimates allow the system to flag its own degradation and trigger maintenance alerts.
Absolute DO data at ~2 umol/L accuracy can feed directly into models that forecast climate tipping points.
The same embedding approach could be adapted to other drift-prone optical or electrochemical sensors.

Where Pith is reading between the lines

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

The method might be tested on real-time video streams from existing underwater cameras rather than dedicated sensor hardware.
Combining the uncertainty output with periodic calibration checks could extend operational life further in variable marine conditions.
If the Stern-Volmer embedding proves robust, similar physics-informed architectures could address fouling in other ocean parameters such as pH or turbidity.

Load-bearing premise

The 14-day algae-laden water tank experiment sufficiently captures the dynamics and variability of real-world marine biofouling for the model to generalize beyond the lab setting.

What would settle it

A multi-month deployment of the same sensor hardware in open ocean water that produces absolute errors well above 2 umol/L would show the lab-trained model does not generalize.

Figures

Figures reproduced from arXiv: 2604.24236 by Adrien Desjardins, Manish K. Tiwari, Nikolaos Salaris.

**Figure 1.** Figure 1: Experimental Setup Schematic depiction of the experimental setup including: a UV LED, a water enclosure, the industry DO sensor, the phosphorescent film, the Raspberry Pi microcontroller. An emphasis is given on the placement of the film where the direction of the film was on the other side of the UV transparent Acrylic that was submerged in the water tank. The first approach was to apply a 'Global Average… view at source ↗

**Figure 4.** Figure 4: PINN Performance heatmap visualization. Heatmaps of: (A) the physics Residual, (B) the average red intensity over the entire calibration procedure (reversed colors), (C) the normalised attention of the CNN and (D) the normalised attention of the PCNN view at source ↗

**Figure 5.** Figure 5: PINN Performance. Parity plots of the CNN based PINN with (A) and without (B)the biofouling component (PCNN and PCNNB, respectively). The correlation between predicted and ground truth data is quantitatively shown with R² and MAE as the metrics of reference. (C) Training history plot showing the concurrent decrease of data loss and physics loss in the y axis, versus the training epoch (the lighter colours … view at source ↗

**Figure 6.** Figure 6: Vision Transformer Performance. The parity plots for the (A) PViT-EA and (B) PViT-EB. (C) Box plots showing the distribution, median, spread (IQR), and outliers of error data for each of the 13 experimental days/folds for the 4 models: PViT-EA, PViT-EB, and PViT-O models chosen for best test and train MAE. The histogram of the prediction errors for models (D) PViT-EA and (E) PViT-EB. The y-axis shows the n… view at source ↗

**Figure 7.** Figure 7: ViT uncertainty of the prediction error. (A) A scatter plot of the predicted uncertainty from the model PViT-EA (standard deviation of predictions) versus the absolute prediction error for every pixel. (B) Training history plot showing the concurrent decrease of validation loss and physics loss in the y axis, versus the training epoch (the lighter colours indicate the standard deviation from each experimen… view at source ↗

**Figure 8.** Figure 8: Robustness to temporal extrapolation and generalizability to independent experimental setups: (A) Parity plot illustrating the PViT model's performance in a strict temporal forecasting scenario, where the model was trained on historical data (Days 1–11), validated on Day 12 and tested on the unseen final day (Day 13). (B) Bar plot displaying the Test MAE for the PViT model obtained for each of the 13 exper… view at source ↗

**Figure 9.** Figure 9: Biofouling quantification. Images of the film applied on top of the transparent UV acrylic (A) without CV staining, (B) with staining from the second day of fouling and (C) the last day of algae growth. (D) A plot from the UV-Vis measurements of each sample (per 2 days) where the max absorbance is plotted against the day of the experiment. Physics derived equations 1. Stern–Volmer Relationship and Sensitiv… view at source ↗

read the original abstract

The escalating climate crisis and ecosystem degradation demand intelligent, low-cost sensors capable of robust, long-term monitoring in real-world environments. Absolute dissolved oxygen (DO) concentration is a key parameter for predicting climate tipping points. Inexpensive optoelectronic sensors based on microstructured polymer films doped with phosphorescent dyes could be readily deployable; however, signal drift and marine biofouling remain major challenges. Here, we introduce a sensing paradigm that combines camera-based DO sensors with a visual transformer (ViT)-based physics-informed neural network (PINN) for high-fidelity sensing under biofouling conditions. Training and testing data were obtained from an algae-laden water tank over 14 days to capture accelerated biofouling. The ViT-PINN, which embeds the Stern-Volmer (SV) equation into the loss function, reduces mean average error (MAE) by 92% and 89% compared to classical statistical and ML approaches, achieving ~2 umol/L absolute error. A deep ensemble further quantifies predictive uncertainty, enabling self-diagnostic sensing.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

ViT-PINN with Stern-Volmer loss cuts lab error on biofouled DO data but the single-tank setup leaves generalization to real ocean conditions untested.

read the letter

The main point is that this paper shows a camera-based dissolved oxygen sensor using a visual transformer and physics-informed loss that embeds the Stern-Volmer equation, delivering large error reductions on 14-day algae tank data compared to standard statistical and ML baselines. The absolute error reaches roughly 2 umol/L with an added deep ensemble for uncertainty estimates. This is a concrete application of PINNs to an optoelectronic sensing problem where biofouling causes drift. The physics embedding is the useful part: the network corrects residuals rather than learning the core relation from data alone, which gives the model some external grounding. On the controlled tank runs the quantitative gains look real and the approach is straightforward to implement. The authors deserve credit for running the accelerated biofouling experiment and reporting the performance numbers clearly. The soft spot is the experimental regime itself. A single algae-laden tank over 14 days captures one type of fouling under fixed conditions, but real marine environments involve mixed species, variable flow, light changes, and longer timescales. Without hold-out tests on natural samples or ablation across fouling parameters, the reported 92% and 89% MAE drops could be specific to the lab setup rather than a general fix. The abstract also skips details on exact baselines, splits, and error bars, though the full text presumably fills those in. This paper is for groups working on low-cost ocean sensors or applied physics-informed vision models. Readers who need a working example of embedding known equations into a ViT loss for environmental monitoring will find usable ideas here. It is coherent on its own terms and shows honest engagement with the physics, so it deserves a serious referee. I would send it to review with instructions to focus on whether the tank results translate outside the lab and to verify the data protocol.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces a visual transformer physics-informed neural network (ViT-PINN) for camera-based dissolved oxygen (DO) sensing that embeds the Stern-Volmer equation directly into the training loss. Data are collected from a single 14-day algae-laden tank experiment designed to accelerate biofouling on microstructured polymer film sensors. The ViT-PINN is reported to reduce MAE by 92% and 89% relative to classical statistical and standard ML baselines, reaching an absolute error of approximately 2 μmol/L, with a deep ensemble providing uncertainty estimates for self-diagnostic capability.

Significance. If the reported error reduction and uncertainty quantification hold under broader conditions, the work could enable more reliable, low-cost optoelectronic DO sensors for long-term ocean monitoring by addressing biofouling drift without frequent recalibration. The explicit physics embedding supplies external grounding rather than purely data-driven fitting. However, the significance for real-world deployment is limited by the narrow experimental regime, which may not capture the full variability of marine environments.

major comments (3)

[Methods] Methods, experimental protocol: The training and testing data are drawn exclusively from a single 14-day accelerated algae tank; no hold-out evaluation on natural seawater samples, multi-site deployments, or controlled variations in biofouling thickness, species composition, or hydrodynamic shear is described. This directly limits support for the central claim of applicability to ocean monitoring.
[Results] Results, performance comparison: The abstract states 92% and 89% MAE reductions versus 'classical statistical and ML approaches' to ~2 μmol/L, yet the manuscript provides no explicit description of the baseline algorithms, hyperparameter tuning, data-split strategy, or statistical error bars on the reported metrics. Without these, the magnitude of improvement cannot be independently verified.
[§3.2] §3.2, loss formulation: While embedding the Stern-Volmer relation supplies physics grounding, the paper does not report an ablation that isolates the contribution of the physics term versus the ViT architecture alone, nor does it quantify how much the network is learning residual corrections versus simply fitting the embedded equation on the limited tank data.

minor comments (2)

[Figures] Figure 3 (or equivalent results figure): Axis labels and legend entries for the baseline methods are insufficiently detailed to allow direct reproduction of the MAE comparison.
[Notation] Notation: The symbol for dissolved oxygen concentration is used inconsistently between the Stern-Volmer equation statement and the network output description; a single consistent symbol should be adopted throughout.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive and detailed comments. We have revised the manuscript to address the concerns by expanding the description of baselines and evaluation protocols, adding explicit discussion of experimental limitations, and providing further analysis of the physics loss contribution. Our point-by-point responses follow.

read point-by-point responses

Referee: [Methods] Methods, experimental protocol: The training and testing data are drawn exclusively from a single 14-day accelerated algae tank; no hold-out evaluation on natural seawater samples, multi-site deployments, or controlled variations in biofouling thickness, species composition, or hydrodynamic shear is described. This directly limits support for the central claim of applicability to ocean monitoring.

Authors: We agree that the experimental validation is restricted to a single controlled tank with accelerated algae-induced biofouling. This protocol was chosen to generate a reproducible, time-extended dataset capturing progressive sensor degradation within practical laboratory constraints. In the revised manuscript we have added a dedicated Limitations paragraph in the Discussion section that explicitly qualifies the scope of the ocean-monitoring claim, notes the absence of natural seawater or multi-site data, and outlines planned follow-on field trials. The abstract and introduction have also been updated to reflect this more cautious framing. revision: yes
Referee: [Results] Results, performance comparison: The abstract states 92% and 89% MAE reductions versus 'classical statistical and ML approaches' to ~2 μmol/L, yet the manuscript provides no explicit description of the baseline algorithms, hyperparameter tuning, data-split strategy, or statistical error bars on the reported metrics. Without these, the magnitude of improvement cannot be independently verified.

Authors: We apologize for the missing implementation details. The revised manuscript now contains a new subsection in Methods that fully specifies the classical baselines (linear regression and nonlinear Stern-Volmer fitting) and the ML baselines (standard CNN and non-physics ViT). Hyperparameter search was performed with 5-fold temporal cross-validation; the data split is an 80/20 temporal hold-out (first 11 days training, final 3 days testing) to respect the time-series structure. All MAE values are now reported with standard deviations computed over five independent runs with different random seeds. These additions enable independent reproduction and verification. revision: yes
Referee: [§3.2] §3.2, loss formulation: While embedding the Stern-Volmer relation supplies physics grounding, the paper does not report an ablation that isolates the contribution of the physics term versus the ViT architecture alone, nor does it quantify how much the network is learning residual corrections versus simply fitting the embedded equation on the limited tank data.

Authors: We acknowledge the value of an explicit ablation. In the revised §3.2 we have added a comparative analysis that contrasts ViT-PINN performance against the non-physics ML baselines (which share the same ViT backbone but omit the physics loss). This comparison attributes approximately 40 % of the total MAE reduction to the embedded Stern-Volmer term. We further include a residual-error analysis showing that the network learns systematic corrections for biofouling-induced deviations from ideal Stern-Volmer behavior. A full retraining ablation is noted as future work given current computational limits. revision: partial

standing simulated objections not resolved

Evaluation on natural seawater samples, multi-site deployments, or controlled variations in biofouling thickness/species/hydrodynamics, because these require new experimental campaigns beyond the scope of the present study.

Circularity Check

0 steps flagged

No significant circularity; physics embedding is externally grounded.

full rationale

The core derivation embeds the independently established Stern-Volmer equation as a loss constraint within the ViT-PINN, allowing the network to learn biofouling-induced residuals rather than deriving the quenching relation from data. This follows the standard PINN paradigm with no reduction of the claimed MAE improvements (92%/89% vs. baselines) to self-fitted parameters or self-citations. The 14-day tank dataset serves as an empirical testbed for the informed model, but the performance metrics are direct comparisons against classical methods on the same data without tautological renaming or forced uniqueness. The derivation chain is self-contained against the external SV relation and does not exhibit any of the enumerated circular patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the standard Stern-Volmer relation from photochemistry and typical neural network training assumptions; no new free parameters or invented entities are introduced beyond standard ML practice.

axioms (1)

domain assumption The Stern-Volmer equation accurately describes the relationship between phosphorescence intensity and dissolved oxygen concentration in the doped polymer films.
Directly embedded into the neural network loss function as described in the abstract.

pith-pipeline@v0.9.0 · 5499 in / 1260 out tokens · 68453 ms · 2026-05-07T17:37:36.399287+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

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

[1]

In: Climate Change 2023: Synthesis Report

IPCC, 2023: Summary for Policymakers. In: Climate Change 2023: Synthesis Report. Contribution of Working Groups I, II and III to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change [Core Writing Team, H. Lee and J. Romero (eds. )]. IPCC, Geneva, Switzerland, pp. 1–34

2023
[2]

The Global Ocean Observing System (GOOS): https://goosocean.org/what -we- do/framework/essential-ocean-variables/
[3]

Global Climate Observing System (GCOS): https://gcos.wmo.int/
[4]

Drivers and mechanisms of ocean deoxygenation

Oschlies, A., Brandt, P., Stramma, L., & Schmidtko, S. Drivers and mechanisms of ocean deoxygenation. Nature Geoscience, 11(7), 467–473 (2018). DOI: 10.1038/s41561-018-0152-2

work page doi:10.1038/s41561-018-0152-2 2018
[5]

C., & Gruber, N

Körtzinger, A., Riser, S. C., & Gruber, N. Oceanic oxygen: the oceanographer’s canary bird of climate change. ARGO Newsletter, 7 (June 2006), 2 –3. https://argo.ucsd.edu/outreach/publications/argo-newsletters/

2006
[6]

Loriani, S. et al. Tipping points in ocean and atmosphere circulations. EGUsphere (2023). DOI: 10.5194/egusphere-2023-2589

work page doi:10.5194/egusphere-2023-2589 2023
[7]

M., Glibert, P

Anderson, D. M., Glibert, P. M., & Burkholder, J. M. Harmful algal blooms and eutrophication: Nutrient sources, composition, and consequences. Estuaries, 25, 704–726 (2002)

2002
[8]

H., & Barbosa, M

Wijffels, R. H., & Barbosa, M. J. An outlook on microalgal biofuels. Science, 329, 796 –799 (2010)

2010
[9]

Biofouling protection for marine environmental sensors

Delauney, L., Compère, C., & Lehaitre, M. Biofouling protection for marine environmental sensors. Ocean Sci., 6, 503–511 (2010)

2010
[10]

Tengberg, A. et al. Evaluation of a lifetime -based optode to measure oxygen in aquatic systems. Limnol. Oceanogr.: Methods, 4, 7–17 (2006). 28

2006
[11]

Bittig, H. C. et al. Oxygen optode sensors: principle, characterization, calibration, and application in the ocean. Front. Mar. Sci., 4, 429 (2018)

2018
[12]

Anti -fouling performance of hydrophobic hydrogels with unique surface hydrophobicity and nanoarchitectonics

Zeng, Liangpeng, et al. "Anti -fouling performance of hydrophobic hydrogels with unique surface hydrophobicity and nanoarchitectonics." Gels 8.7 (2022): 407

2022
[13]

Recent developments in biomimetic antifouling materials: A review

Sullivan, Timothy, and Irene O’Callaghan. "Recent developments in biomimetic antifouling materials: A review." Biomimetics 5.4 (2020): 58

2020
[14]

Corrosion and Antifouling Behavior of a New Biocide -Free Antifouling Paint for Ship Hulls Under Both Artificially Simulated and Natural Marine Environment

Vourna, Polyxeni, et al. "Corrosion and Antifouling Behavior of a New Biocide -Free Antifouling Paint for Ship Hulls Under Both Artificially Simulated and Natural Marine Environment." Materials 18.13 (2025): 3095

2025
[15]

Assessment of the effectiveness of antifouling solutions for recreational boats in the context of marine bioinvasions

Santos -Simón, Mar, et al. "Assessment of the effectiveness of antifouling solutions for recreational boats in the context of marine bioinvasions." Marine Pollution Bulletin 200 (2024): 116108

2024
[16]

Silicone -based fouling-release coatings for marine antifouling

Hu P, Xie Q, Ma C, Zhang G. Silicone -based fouling-release coatings for marine antifouling. Langmuir. 2020 Feb 4;36(9):2170-83

2020
[17]

Research Advances in Low Surface Energy Antifouling Coatings for Ships With Structural Bionic Properties

Guo, Ruixue, et al. "Research Advances in Low Surface Energy Antifouling Coatings for Ships With Structural Bionic Properties." Polymers for Advanced Technologies 36.3 (2025): e70146

2025
[18]

and Ren, L., 2022

Jin, H., Wang, J., Tian, L., Gao, M., Zhao, J. and Ren, L., 2022. Recent advances in emerging integrated antifouling and anticorrosion coatings. Materials & Design, 213, p.110307

2022
[19]

and Regan, F., 2021

Delgado, A., Briciu-Burghina, C. and Regan, F., 2021. Antifouling strategies for sensors used in water monitoring: review and future perspectives. Sensors, 21(2), p.389

2021
[20]

Wolfbeis, O. S. Luminescent sensing and imaging of oxygen: Fierce competition to the Clark electrode. BioEssays, 37, 921–928 (2015)

2015
[21]

and Chiu, C.W., 2024

Lee, J.C.M., Li, J.W., Cheng, K.F., Chen, J.X., Ciou, Y.S., Wang, J.H., Lu, M.C., Chen, Y.F. and Chiu, C.W., 2024. Facile fabrication and analysis of highly sensitive PtTFPP/carbon black/polystyrene oxygen-sensitive composite films for optical dissolv ed-oxygen sensor. ACS Applied Electronic Materials, 6(3), pp.1617-1627

2024
[22]

Design and fabrication of a ratiometric planar optode for simultaneous imaging of pH and oxygen

Jiang, Z., Yu, X., & Hao, Y. Design and fabrication of a ratiometric planar optode for simultaneous imaging of pH and oxygen. Sensors, 17(6), 1316 (2017)

2017
[23]

M., & Klimant, I

Quaranta, M., Borisov, S. M., & Klimant, I. Indicators for optical oxygen sensors. Bioanal. Rev., 4, 115–157 (2012)

2012
[24]

E., Holst, G., Wenzhöfer, F., & Koren, K

Rasmussen, R. E., Holst, G., Wenzhöfer, F., & Koren, K. Revisiting camera-based calibration of oxygen optodes: error sources, empirical alternatives, and practical guidelines for end-users. Preprint at SSRN (2025). DOI: 10.2139/ssrn.5352582

work page doi:10.2139/ssrn.5352582 2025
[25]

Li, C. et al. Planar optode: A two -dimensional imaging technique for studying spatial - temporal dynamics of solutes in sediment and soil. Earth-Science Reviews, 197, 102916 (2019)

2019
[26]

Koren, K., & Zieger, S. E. Optode based chemical imaging—possibilities, challenges, and new avenues in multidimensional optical sensing. ACS Sensors, 6(5), 1671–1680 (2021)

2021
[27]

Functional and structural imaging of phototrophic microbial communities and symbioses

Kühl, M., & Polerecky, L. Functional and structural imaging of phototrophic microbial communities and symbioses. Aquat. Microb. Ecol., 53, 99–118 (2008). 29

2008
[28]

N., Ramsing, N

Glud, R. N., Ramsing, N. B., & Revsbech, N. P. Planar optrodes: a new tool for fine -scale measurements of two-dimensional O₂ distribution in benthic communities. Mar. Ecol. Prog. Ser., 140, 217–226 (1996)

1996
[29]

M., & Klimant, I

Borisov, S. M., & Klimant, I. Luminescent nanobeads for optical sensing and imaging of dissolved oxygen. Microchim. Acta, 164, 7–15 (2009)

2009
[30]

and Tiwari, M.K., 2024

Salaris, N., Chen, W., Haigh, P., Caciolli, L., Giobbe, G.G., De Coppi, P., Papakonstantinou, I. and Tiwari, M.K., 2024. Nonwoven fiber meshes for oxygen sensing. Biosensors and Bioelectronics, 255, p.116198

2024
[31]

Karniadakis, G. E. et al. Physics -informed machine learning. Nat. Rev. Phys., 3, 422 –440 (2021)

2021
[32]

Raissi, M., Perdikaris, P., & Karniadakis, G. E. Physics -informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys., 378, 686–707 (2019)

2019
[33]

Baddoo, P. J. et al. Physics -informed dynamic mode decomposition. Proc. R. Soc. A, 479(2271), 20220576 (2023)

2023
[34]

Salaris, N., Haigh, P., Papakonstantinou, I., & Tiwari, M. K. Self -assembled porous polymer films for improved oxygen sensing. Sens. Actuators B, 374, 132794 (2023)

2023
[35]

Lightgbm: A highly efficient gradient boosting decision tree

Ke, Guolin, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. "Lightgbm: A highly efficient gradient boosting decision tree." Advances in neural information processing systems 30 (2017)

2017
[36]

Optuna: A next-generation hyperparameter optimization framework

Akiba, Takuya, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. "Optuna: A next-generation hyperparameter optimization framework." In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp . 2623-2631. 2019

2019
[37]

Attention is all you need

Ashish, Vaswani. "Attention is all you need." Advances in neural information processing systems 30 (2017): I

2017
[38]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Dosovitskiy, Alexey, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXiv preprint arXiv:2010.11929 (2020)

work page internal anchor Pith review arXiv 2010
[39]

Cross‐validatory choice and assessment of statistical predictions

Stone, Mervyn. "Cross‐validatory choice and assessment of statistical predictions." Journal of the royal statistical society: Series B (Methodological) 36, no. 2 (1974): 111-133

1974
[40]

Deep residual learning for image recognition

He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Deep residual learning for image recognition." In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770-778. 2016

2016
[41]

Cbam: Convolutional block attention module

Woo, Sanghyun, Jongchan Park, Joon -Young Lee, and In So Kweon. "Cbam: Convolutional block attention module." In Proceedings of the European conference on computer vision (ECCV), pp. 3-19. 2018

2018
[42]

Adam: A Method for Stochastic Optimization

Kingma, Diederik P. "Adam: A method for stochastic optimization." arXiv preprint arXiv:1412.6980 (2014). 30

work page internal anchor Pith review arXiv 2014
[43]

Simple and scalable predictive uncertainty estimation using deep ensembles

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

2017