Pith. sign in

REVIEW 3 major objections 6 minor 49 references

FastCAR: Fast Classification And Regression for Task Consolidation in Multi-Task Learning to Model a Continuous Property Variable of Detected Object Class

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single regression network with interval-encoded labels solves both classification and property regression that multi-task models fail on.

desk verdict A simple label-shift trick that plausibly works on a new steel dataset, but the headline MTL comparison is undermined by an unfair baseline. read the letter →

arxiv 2506.00208 v1 pith:ZI2R4KVD submitted 2025-05-30 cs.CV

Anoop Kini , Andreas Jansche , Timo Bernthaler , Gerhard Schneider This is my paper
classification cs.CV
keywords multi-tasklearningtaskconsolidationlabeltransformationclassificationandregressionpropertymodelingsteelmicrostructurehardnesspredictionsingle-tasknetwork
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 claims that a single-task regression network, trained on labels that encode both object class and a continuous property value in one number, can solve both a classification and a regression task that standard multi-task networks fail to learn together. On a new dataset of 4,536 steel micrograph images with six classes and continuous hardness labels, the method, called FastCAR, reaches 99.54% classification accuracy and a regression mean absolute percentage error of 2.4%. It trains 2.52x quicker and infers 55% faster than the best-performing benchmark multi-task model, which achieved comparable classification accuracy but a regression error of about 60 K. The significance is that task grouping for weakly correlated heterogeneous tasks may not need complex architectures or loss-weighting schemes; a label transformation can consolidate the tasks.

What carries the argument

The central object is the transformed label $g_i(x) = f_i(x) + k_i$, where $f_i(x)$ is the property value (regression label) for an instance of class $C_i$ and $k_i = (i-1) \cdot u \cdot \delta$, with $\delta$ the maximum class-wise property range and $u$ chosen in $[1,2]$. The offsets place the class intervals $S_i$ in disjoint, sorted, well-separated positions, so a single scalar output can be decoded into a class index and a property value. A final centering step subtracts the midpoint of the union of intervals to keep the regression targets near zero. This label transformation is what lets a plain ResNet-18 regression network with one output neuron learn both tasks, avoiding separate classification and regression heads and any explicit loss balancing.

What would settle it

Measure the six class-wise hardness ranges in the Advanced Steel Property Dataset and compute the shifted intervals from Algorithm 1; if any pair of shifted intervals overlaps by more than the tolerance implied by $u$, classification accuracy on those classes should collapse to chance. Equivalently, run FastCAR on data whose adjacent class property ranges overlap by more than $(2-u)\cdot\delta$ and verify that classification accuracy drops well below the reported 99.54%.

Watch

Extended reading notes

Core claim

FastCAR establishes that a classification task and a regression task with only subtle correlation can be consolidated into one regression task by shifting each class's property interval to a distinct, ordered, well-separated location on the real line. The class index is read off from which interval the predicted value falls into, and the property value is recovered by subtracting the class-specific offset. On the contributed Advanced Steel Property Dataset, this single scalar-label scheme outperforms 252 benchmark multi-task models on the joint objective: 99.54% classification accuracy and roughly 0.5 K MSE (2.4% MAPE), while the best MTL baseline (Cross-stitch with IMTL weighting) reached 99.69% accuracy but 60 K MSE (over 55% MAPE). The paper argues that existing MTL architectures and weighting schemes cannot jointly handle the two tasks on this data, whereas the label transformation absorbs the classification structure directly into the regression target.

Load-bearing premise

The whole scheme relies on the property ranges of different classes being separated enough that after shifting they do not overlap; if two classes have overlapping property intervals, the single scalar label cannot be uniquely decoded into a class.

Editorial extensions

If this is right

  • A single-task regression network can serve as a drop-in replacement for a multi-task network in property-modeling problems where class-wise property ranges are sufficiently separated, reducing model size and inference cost.
  • The label transformation implicitly handles gradient feedback between classification and regression, removing the need for loss-weighting schemes such as GradNorm, IMTL, or uncertainty weighting.
  • Because the method only re-labels the targets, it can be combined with any architecture and any regression loss, so improvements to single-task regression directly transfer to the joint task.
  • Benchmarking MTL methods on this dataset shows that 252 architecture/weighting combinations fail to learn both tasks, indicating that task consolidation via label engineering is a viable alternative when shared representations are weak or absent.

Reading between the lines

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

  • If the separation condition holds, FastCAR should generalize to other domains where a continuous property is measured per class, such as temperature estimation of thermal cloaks or mechanical property prediction from micrographs; the paper lists these as related but does not test them.
  • A stress test would deliberately use classes whose property intervals overlap after shifting by more than $(2-u)\cdot\delta$; the paper gives no evidence about how gracefully classification degrades as intervals approach overlap, so the method's practical envelope is unknown.
  • One could extend the encoding to overlapping intervals by using a two-dimensional target or a soft interval membership, but that would reintroduce a second prediction head and lose the single-output property, so the trade-off merits explicit study.
  • The reported 2.52x training speedup and 55% inference speedup are measured against the best-performing MTL baseline; against simpler baselines like CGC with GradVac the speedup is larger, but the comparison set does not include a pure single-task classifier or regressor trained separately.
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

3 major / 6 minor

Summary. The paper proposes FastCAR, a label-transformation method that encodes both a class label and a continuous property value into a single scalar target, so that a plain regression network can perform classification and regression jointly. On a new 'Advanced Steel Property Dataset' of 4536 optical microscopy images with six microstructure classes and hardness as the continuous property, FastCAR is reported to reach 99.54% classification accuracy and about 2.4% MAPE for hardness regression, while benchmark MTL models (96 or 252 configurations from the architecture/weighting landscape) are reported to fail at regression. The paper further claims 2.52x faster training and 55% faster inference than the best-performing MTL model.

Significance. The core idea is simple and attractive: if the property ranges of different classes are separable, the two tasks can be consolidated into one regression target, which may be useful for many property-modeling applications. The released dataset is a new contribution to a domain with few public benchmarks. The method's feasibility on this dataset is plausible. However, the paper leaves a key data-dependent condition (disjointness and separation of class intervals) unverified, and its comparative claim against MTL baselines rests on a regression failure that is plausibly an artifact of loss-scale imbalance rather than an intrinsic property of MTL. The contribution is therefore potentially useful but, as presented, not fully established.

major comments (3)
  1. [Sec. 3.2, Algorithm 1, Eqs. (2a)-(2c)] The transformation's correctness hinges on the original class-wise property ranges being sufficiently separated. The paper asserts the disjointness, ordering, and separation conditions but does not prove them from the data or even report the per-class intervals [ai,bi]. In fact, the gap between shifted intervals is (a_{i+1}-b_i)+u*delta, which is only guaranteed to lie in (delta, 2*delta) under strong assumptions (e.g., equal-length contiguous ranges). Since the ablation in Table 1 shows that 'Bad' transformed labels collapse classification accuracy to about 19%, the condition is empirically load-bearing. The authors should report the actual class ranges, verify that Eqs. (2a)-(2c) hold for the used u, or state the explicit data condition required for correctness.
  2. [Sec. 4.1, Sec. 4.2, Sec. 6.2] The claim that FastCAR outperforms all benchmark MTL models rests entirely on the observation that no MTL model learns the regression task (best MAPE >54%). The paper itself concedes in Sec. 4.2 that 'a mismatch of the relative scale of the weights between classification and regression tasks is possible.' Since cross-entropy is order-1 while hardness MSE is in units of squared Kelvin (order 10^2-10^4), the regression gradient can be swamped. No target standardization (e.g., zero-mean unit-variance normalization) and no systematic per-task loss-weight grid was performed for the MTL baselines; Sec. 6.2 reports hyperparameter tuning only for MGDA and STCH. Given that the same ResNet-18 backbone as a single-task regressor achieves MAPE 2.4%, the data and features do support regression, so the universal MTL failure is not demonstrated to be intrinsic. The comparative superiority claim requires either properly balanced MTL baselines or a more modest claim scoped to the tested configuration.
  3. [Sec. 4.3, Table 1] The ablation states that 'a prediction layer with one neuron (a regression network)' is critical, but the rows with >1 neurons are marked N/A, so no alternative is actually tested. The reader cannot tell whether a multi-neuron output layer combined with 'Good' labels would also succeed or fail. Please either run those conditions or rephrase the conclusion to avoid implying a controlled comparison.
minor comments (6)
  1. [Abstract, Sec. 4.1, Sec. 5, Table 1] The MAPE is reported as 2.4% in the abstract and Sec. 4.1, but Table 1 shows 2.6% with range centering and 2.4% without; the conclusion also states a classification accuracy of 99.45% instead of 99.54%. Please reconcile these numbers.
  2. [Eqs. (2b), (2c)] The notation S1 < S2 < ... < Sn and min_i S_{i+1}, max_i S_i for intervals is not defined. Please define an ordering on closed intervals and specify whether the min/max are over elements of the interval or over the index i.
  3. [Table 1] The column header 'ki (Good / Bad)' is confusing; the transformed label is gi, not ki. Also, it is not specified how the 'Bad' labels were generated -- indicate e.g. which u values were used and why the resulting labels violate Eqs. (2a)-(2c).
  4. [Sec. 6.5] The sentence 'Hence, a direct comparison with FastCAR is not feasible' appears to contradict the benchmarking results presented in the paper. Please clarify what was not feasible (e.g., larger architectural variants exceeded GPU memory) and how this affects the reported comparison.
  5. [Sec. 4.1] The statement that MTL models 'can learn the classification task in about half of the cases' would benefit from a quantitative criterion (e.g., accuracy above a threshold) so the reader can interpret Fig. 4.
  6. [Sec. 3.2] The text states the gap between consecutive shifted intervals 'lies between δ and 2*δ' without derivation. Please add a derivation or note the data-dependent nature of this claim (see major comment 1).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the label transformation is a deliberate input encoding, and the reported accuracy and error are measured on held-out data, not forced by construction.

full rationale

The paper's derivation chain is: (1) define each class's property interval Ri = [ai, bi]; (2) shift these intervals by class-dependent offsets ki = (i-1)*u*delta with u in [1,2]; (3) train a single-task regression network on the transformed labels gi(x) = fi(x) + ki; (4) decode the class index by locating the predicted transformed value in the shifted interval Si. The transformed labels are constructed from the training data's own regression ranges, but this is the method being proposed, not a hidden fit: the reported classification accuracy (99.54%) and regression MAPE (2.4%) are measured on held-out test data. The ablation study makes the non-triviality explicit: with 'Bad' labels that do not satisfy Eqs. (2a)-(2c), classification accuracy collapses to 18.82% even though regression MSE is lower, so the classification accuracy is not a tautological consequence of the interval encoding. The empirical guideline in Sec. 6.1 (validation-loss and gradient thresholds) is an early-stopping/health check, not a parameter fitted to the reported metrics. There are no load-bearing self-citations: the dataset is self-contributed, but that is a data resource rather than a premise used to derive the results. The main scientific weakness is that the MTL comparison may be unfair because per-task loss scales were not swept, and the paper itself concedes a possible scale mismatch; however, that is an external-validity and benchmarking concern, not circularity in the derivation chain.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method relies on the separability of class-wise property ranges, a domain assumption not guaranteed by the algorithm, plus standard assumptions about network training and fair benchmarking. No new physical entities are introduced; the transformed label is a mathematical re-encoding of existing labels.

free parameters (1)
  • u (offset multiplier) = in [1,2], exact value not reported
    Hyperparameter controlling gap size between class intervals after shifting. Affects classification margin and gradient behavior; not tuned or reported, and no sensitivity analysis is given.
assumptions (3)
  • domain assumption Class-wise property ranges can be made disjoint by adding offsets k_i = (i-1)*u*delta with u in [1,2].
    Algorithm 1 sets offsets solely from delta and u; it does not verify or guarantee Eqs. (2a)-(2c) for arbitrary input ranges. If adjacent class ranges overlap substantially, no u <= 2 separates them.
  • domain assumption A single-neuron regression network can learn the transformed label well enough that interval decoding yields both class and property.
    The method relies on the network's regression accuracy on the artificial target; no theoretical or empirical guarantee is given beyond the reported single run.
  • domain assumption Benchmark MTL models were configured and tuned fairly.
    The claim of outperforming all MTL baselines depends on the 252-model benchmark being representative; the universal regression failure of MTL models suggests possible scaling or tuning issues, and only MGDA and STCH received extra hyperparameter tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FastCAR: Fast Classification And Regression for Task Consolidation in Multi-Task Learning to Model a Continuous Property Variable of Detected Object Class." pith.science (2026). https://pith.science/paper/ZI2R4KVD

@misc{pith2026250600208,
  author       = {Pith},
  title        = {Pith review of: FastCAR: Fast Classification And Regression for Task Consolidation in Multi-Task Learning to Model a Continuous Property Variable of Detected Object Class},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZI2R4KVD}},
  note         = {Machine review of arXiv:2506.00208}
}
read the original abstract

FastCAR is a novel task consolidation approach in Multi-Task Learning (MTL) for a classification and a regression task, despite the non-triviality of task heterogeneity with only a subtle correlation. The approach addresses the classification of a detected object (occupying the entire image frame) and regression for modeling a continuous property variable (for instances of an object class), a crucial use case in science and engineering. FastCAR involves a label transformation approach that is amenable for use with only a single-task regression network architecture. FastCAR outperforms traditional MTL model families, parametrized in the landscape of architecture and loss weighting schemes, when learning both tasks are collectively considered (classification accuracy of 99.54%, regression mean absolute percentage error of 2.4%). The experiments performed used "Advanced Steel Property Dataset" contributed by us https://github.com/fastcandr/AdvancedSteel-Property-Dataset. The dataset comprises 4536 images of 224x224 pixels, annotated with discrete object classes and its hardness property that can take continuous values. Our proposed FastCAR approach for task consolidation achieves training time efficiency (2.52x quicker) and reduced inference latency (55% faster) than benchmark MTL networks.

Figures

Figures reproduced from arXiv: 2506.00208 by the authors.

Figure 1
Figure 1. This is a schematic representation of the FastCAR ap [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic representation of regression labels and transformed labels (or hybrid labels) in a class-wise manner using a box plot [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Decoder architecture of benchmark multi-task learning [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance of multi-task learning (MTL) models on the classification task (in terms of classification accuracy in %) across [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Performance of multi-task learning models on the regression task (in terms of MSE error expressed in multiples of 1000) across [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 41 canonical work pages

  1. [1]

    Machine learning for predicting properties of porous media from 2d x-ray images

    Naif Alqahtani, Fatimah Alzubaidi, Ryan T Armstrong, Pawel Swietojanski, and Peyman Mostaghimi. Machine learning for predicting properties of porous media from 2d x-ray images. Journal of Petroleum Science and Engineer- ing, 184:106514, 2020. 3

    work page 2020

  2. [2]

    Digitization and image-based structure-properties relationship evaluation of a porous gold micro-electrode

    Anthony Baux, Guillaume Cou ´egnat, G ´erard L Vignoles, Didier Lasseux, Alexander Kuhn, Cristina Carucci, Nico- las Mano, and Tien Dung Le. Digitization and image-based structure-properties relationship evaluation of a porous gold micro-electrode. Materials & Design, 193:108812, 2020. 3

    work page 2020

  3. [3]

    Multitask learning: A knowledge-based source of inductive bias1

    R Caruana. Multitask learning: A knowledge-based source of inductive bias1. In Proceedings of the Tenth International Conference on Machine Learning , pages 41–48. Citeseer,

    work page

  4. [4]

    Predicting the temperature field of thermal cloaks in homogeneous isotropic multilayer materials based on deep learning

    Haolong Chen, Xinyue Tang, Zhaotao Liu, Zhanli Liu, and Huanlin Zhou. Predicting the temperature field of thermal cloaks in homogeneous isotropic multilayer materials based on deep learning. International Journal of Heat and Mass Transfer, 219:124849, 2024. 3

    work page 2024

  5. [5]

    Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and An- drew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In In- ternational conference on machine learning, pages 794–803. PMLR, 2018. 2, 3, 7

    work page 2018

  6. [6]

    Just pick a sign: Optimizing deep multitask models with gra- dient sign dropout

    Zhao Chen, Jiquan Ngiam, Yanping Huang, Thang Luong, Henrik Kretzschmar, Yuning Chai, and Dragomir Anguelov. Just pick a sign: Optimizing deep multitask models with gra- dient sign dropout. Advances in Neural Information Process- ing Systems, 33:2039–2050, 2020. 3

    work page 2020

  7. [7]

    Multinet++: Multi-stream feature ag- gregation and geometric loss strategy for multi-task learning

    Sumanth Chennupati, Ganesh Sistu, Senthil Yogamani, and Samir A Rawashdeh. Multinet++: Multi-stream feature ag- gregation and geometric loss strategy for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3, 7

    work page 2019

  8. [8]

    Multi-task deep learning as multi-objective op- timization

    Giuseppe Di Fatta, Giuseppe Nicosia, Varun Ojha, and Panos Pardalos. Multi-task deep learning as multi-objective op- timization. In Encyclopedia of Optimization , pages 1–10. Springer, 2023. 1

    work page 2023

Show all 49 references
  1. [9]

    Improving image derived vegetation maps with regression based distribution modeling

    SZ Dobrowski, JA Greenberg, CM Ramirez, and SL Ustin. Improving image derived vegetation maps with regression based distribution modeling. Ecological Modelling, 192(1- 2):126–142, 2006. 3

    2006

  2. [10]

    Dermatologist-level classification of skin cancer with deep neural networks

    Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. nature, 542(7639):115–118, 2017. 3

    2017

  3. [11]

    Tood: Task-aligned one-stage object detec- tion

    Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R Scott, and Weilin Huang. Tood: Task-aligned one-stage object detec- tion. In 2021 IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 3490–3499. IEEE Computer So- ciety, 2021. 3

    2021

  4. [12]

    Miti- gating gradient bias in multi-objective learning: A provably convergent approach

    Heshan Devaka Fernando, Han Shen, Miao Liu, Subhajit Chaudhury, Keerthiram Murugesan, and Tianyi Chen. Miti- gating gradient bias in multi-objective learning: A provably convergent approach. In The Eleventh International Confer- ence on Learning Representations, 2022. 3, 7

    2022

  5. [13]

    Diagnostic ability of heidelberg retina tomograph 3 classifications: glau- coma probability score versus moorfields regression analy- sis

    Antonio Ferreras, Ana B Pajar ´ın, Vicente Polo, Jos´e M Lar- rosa, Lu´ıs E Pablo, and Francisco M Honrubia. Diagnostic ability of heidelberg retina tomograph 3 classifications: glau- coma probability score versus moorfields regression analy- sis. Ophthalmology, 114(11):1981–1...

    1981

  6. [14]

    Efficiently identifying task groupings for multi-task learning

    Chris Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. Advances in Neural Information Pro- cessing Systems, 34:27503–27516, 2021. 3

    2021

  7. [15]

    Ota: Optimal transport assignment for object detection

    Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, and Jian Sun. Ota: Optimal transport assignment for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 303–312, 2021. 3

    2021

  8. [16]

    Learn- ing to branch for multi-task learning

    Pengsheng Guo, Chen-Yu Lee, and Daniel Ulbricht. Learn- ing to branch for multi-task learning. InInternational confer- ence on machine learning, pages 3854–3863. PMLR, 2020. 1, 2, 3, 7, 8

    2020

  9. [17]

    A multi-task learning-based optimiza- tion approach for finding diverse sets of microstructures with desired properties

    Tarek Iraki, Lukas Morand, Johannes Dornheim, Norbert Link, and Dirk Helm. A multi-task learning-based optimiza- tion approach for finding diverse sets of microstructures with desired properties. Journal of Intelligent Manufacturing, 35 (4):1887–1903, 2024. 3

    1903

  10. [18]

    Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geome- try and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482–7491,

  11. [19]

    Predicting the effective me- chanical property of heterogeneous materials by image based modeling and deep learning

    Xiang Li, Zhanli Liu, Shaoqing Cui, Chengcheng Luo, Chenfeng Li, and Zhuo Zhuang. Predicting the effective me- chanical property of heterogeneous materials by image based modeling and deep learning. Computer Methods in Applied Mechanics and Engineering, 347:735–753, 2019. 3

    2019

  12. [20]

    Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection

    Xiang Li, Wenhai Wang, Lijun Wu, Shuo Chen, Xiaolin Hu, Jun Li, Jinhui Tang, and Jian Yang. Generalized focal loss: Learning qualified and distributed bounding boxes for dense object detection. Advances in Neural Information Processing Systems, 33:21002–21012, 2020. 2, 3

    2020

  13. [21]

    Libmtl: A python library for deep multi-task learning

    Baijiong Lin and Yu Zhang. Libmtl: A python library for deep multi-task learning. Journal of Machine Learning Re- search, 24(1-7):18, 2023. 3, 6, 7

    2023

  14. [22]

    Reasonable effectiveness of random weighting: A litmus test for multi-task learning

    Baijiong Lin, Feiyang Ye, Yu Zhang, and Ivor W Tsang. Reasonable effectiveness of random weighting: A litmus test for multi-task learning. arXiv preprint arXiv:2111.10603 ,

    arXiv

  15. [23]

    Dual-balancing for multi-task learning

    Baijiong Lin, Weisen Jiang, Feiyang Ye, Yu Zhang, Peng- guang Chen, Ying-Cong Chen, Shu Liu, and James T Kwok. Dual-balancing for multi-task learning. arXiv preprint arXiv:2308.12029, 2023. 3

    2023

  16. [24]

    Smooth tchebycheff scalar- ization for multi-objective optimization

    Xi Lin, Xiaoyuan Zhang, Zhiyuan Yang, Fei Liu, Zhenkun Wang, and Qingfu Zhang. Smooth tchebycheff scalar- ization for multi-objective optimization. arXiv preprint arXiv:2402.19078, 2024. 3, 7

    2024 arXiv

  17. [25]

    Conflict-averse gradient descent for multi-task learn- ing

    Bo Liu, Xingchao Liu, Xiaojie Jin, Peter Stone, and Qiang Liu. Conflict-averse gradient descent for multi-task learn- ing. Advances in Neural Information Processing Systems , 34:18878–18890, 2021. 3 9

    2021

  18. [26]

    Towards impartial multi-task learning

    Liyang Liu, Yi Li, Zhanghui Kuang, J Xue, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Towards impartial multi-task learning. In Towards impartial multi- task learning. iclr, 2021. 3

    2021

  19. [27]

    End- to-end multi-task learning with attention

    Shikun Liu, Edward Johns, and Andrew J Davison. End- to-end multi-task learning with attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1871–1880, 2019. 1, 2, 3, 7

    2019

  20. [28]

    Asm micrograph database™

    Afina Lupulescu, Tyler Flowers, Linda Vermillion, and Scott Henry. Asm micrograph database™. Metallography, Mi- crostructure, and Analysis, 4:322–327, 2015. 5

    2015

  21. [29]

    Modeling task relationships in multi-task learning with multi-gate mixture-of-experts

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceed- ings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining , pages 1930–1939,

    1930

  22. [30]

    An overview of microstruc- tural and material properties of ultra-high-performance con- crete

    Onkar Mishra and SP Singh. An overview of microstruc- tural and material properties of ultra-high-performance con- crete. Journal of Sustainable Cement-Based Materials, 8(2): 97–143, 2019. 3

    2019

  23. [31]

    Cross-stitch networks for multi-task learning

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Mar- tial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994–4003, 2016. 1, 2, 3, 7

    2016

  24. [32]

    Electro- formed iron as new biomaterial for degradable stents: Devel- opment process and structure–properties relationship

    M Moravej, F Prima, M Fiset, and D Mantovani. Electro- formed iron as new biomaterial for degradable stents: Devel- opment process and structure–properties relationship. Acta biomaterialia, 6(5):1726–1735, 2010. 3

    2010

  25. [33]

    Multi- task learning as a bargaining game

    Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multi- task learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 3, 7

    2022 arXiv

  26. [34]

    Joint classification and trajectory regression of online handwriting using a multi-task learn- ing approach

    Felix Ott, David R ¨ugamer, Lucas Heublein, Bernd Bischl, and Christopher Mutschler. Joint classification and trajectory regression of online handwriting using a multi-task learn- ing approach. In Proceedings of the IEEE/CVF winter con- ference on applications of computer visi...

  27. [35]

    Testing multivariate adaptive regression splines (mars) as a method of land cover classification of terra-aster satellite im- ages

    Elia Quir ´os, ´Angel M Felic ´ısimo, and Aurora Cuartero. Testing multivariate adaptive regression splines (mars) as a method of land cover classification of terra-aster satellite im- ages. Sensors, 9(11):9011–9028, 2009. 3

    2009

  28. [36]

    Low-latency task-oriented communications with multi-round, multi-task deep learning

    Yalin E Sagduyu, Tugba Erpek, Aylin Yener, and Sennur Ulukus. Low-latency task-oriented communications with multi-round, multi-task deep learning. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 2365–2370, 2024. 1, 8

    2024

  29. [37]

    Multi-task learning as multi-objective optimization

    Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization. Advances in neural informa- tion processing systems, 31, 2018. 1, 3, 7

    2018

  30. [38]

    Independent component alignment for multi-task learning

    Dmitry Senushkin, Nikolay Patakin, Arseny Kuznetsov, and Anton Konushin. Independent component alignment for multi-task learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20083–20093, 2023. 3, 7

    2023

  31. [39]

    Which tasks should be learned together in multi-task learning? In International Conference on Machine Learning, pages 9120–9132

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. Which tasks should be learned together in multi-task learning? In International Conference on Machine Learning, pages 9120–9132. PMLR,

  32. [40]

    Mechanical response of human red blood cells in health and disease: Some structure-property-function rela- tionships

    Subra Suresh. Mechanical response of human red blood cells in health and disease: Some structure-property-function rela- tionships. Journal of materials research, 21(8):1871–1877,

  33. [41]

    Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 269–278, 2020. 1, 2, 3, 7

    2020

  34. [42]

    A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas

    Juan Terven, Diana-Margarita C ´ordova-Esparza, and Julio- Alejandro Romero-Gonz ´alez. A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas. Machine Learning and Knowledge Ex- traction, 5(4):1680–1716, 2023. 2

    2023

  35. [43]

    Evo- lutionary neural architecture search for image restoration

    Gerard Jacques van Wyk and Anna Sergeevna Bosman. Evo- lutionary neural architecture search for image restoration. In 2019 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2019. 1

    2019

  36. [44]

    Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models

    Zirui Wang, Yulia Tsvetkov, Orhan Firat, and Yuan Cao. Gradient vaccine: Investigating and improving multi-task optimization in massively multilingual models. arXiv preprint arXiv:2010.05874, 2020. 3, 7

    2010 arXiv

  37. [45]

    Heterogeneous multitask learning with joint sparsity constraints

    Xiaolin Yang, Seyoung Kim, and Eric Xing. Heterogeneous multitask learning with joint sparsity constraints. Advances in neural information processing systems, 22, 2009. 1

    2009

  38. [46]

    Gradient surgery for multi-task learning

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Pro- cessing Systems, 33:5824–5836, 2020. 3, 7

    2020

  39. [47]

    A survey on multi-task learning

    Yu Zhang and Qiang Yang. A survey on multi-task learning. IEEE Transactions on Knowledge and Data Engineering, 34 (12):5586–5609, 2021. 1

    2021

  40. [48]

    Facial landmark detection by deep multi-task learning

    Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xiaoou Tang. Facial landmark detection by deep multi-task learning. In Computer Vision–ECCV 2014: 13th European Confer- ence, Zurich, Switzerland, September 6-12, 2014, Proceed- ings, Part VI 13, pages 94–108. Springer, 2014. 2, 3...

    2014

  41. [49]

    Supplementary section 6.1. Empirical guideline for FastCAR algorithm Following empirically determined guideline was used for assessing the progress of FastCAR algorithm for learning both tasks, in terms of validation loss and average gradient across trainable parameters of the...

Pith tools

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