Pith. sign in

REVIEW 4 major objections 5 minor 47 references

By blocking gradients between hierarchical tasks, FISHER raises ultra-rare fish species accuracy to 63.8%, 13.4 points over a strong baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 07:53 UTC pith:57GCBNON

load-bearing objection Gradient detachment is a sensible design for a biologically structured MTL hierarchy, but the decoupling-specific gains are small single-run margins; the headline 13.4-point ultra-rare gain is mostly auxiliary-task supervision. the 4 major comments →

arxiv 2607.20523 v1 pith:57GCBNON submitted 2026-07-09 q-bio.QM eess.IV

FISHER: Gradient-Decoupled Hierarchical Multi-Task Learning for Fine-Grained Aquatic Species Recognition

Phuc H. Nguyen , Ba Hung Ngo , Mai Phuong Tran , Cuong D. Do , Van-Dinh Nguyen This is my paper
classification q-bio.QM eess.IV
keywords fine-grained aquatic species recognitionmulti-task learninggradient decouplingnegative transferhierarchical task formulationmorphological traitsprototype-based segmentationlong-tailed classification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that gradient interference between low-level segmentation and high-level species classification is the main cause of negative transfer in fine-grained fish recognition. It proposes FISHER, a multi-task architecture that forces a one-way flow of information — from anatomical segmentation to morphological traits to species identity — and stops gradients at each boundary so that classification cannot corrupt the features segmentation learns. The central empirical result is that this design lifts accuracy on ultra-rare species (fewer than ten training images) from 50.4% to 63.8%, and raises trait prediction on unseen species to 97.7% mean average precision. If right, the paper shows that respecting the biological hierarchy of traits, rather than simply training all tasks in parallel, is what lets rare classes inherit knowledge from shared anatomical cues.

Core claim

FISHER's core claim is that gradient conflicts between dense segmentation and global classification are a fundamental source of negative transfer, and that explicitly decoupling gradients across a biologically ordered task chain fixes it. The architecture factorizes the joint distribution as P(y,t,M|X)=P(y|t,X)P(t|M,X)P(M|X), and makes each factor a separate stage: a prototype-based segmentation head produces part masks, a trait head reads masked part evidence plus global features, and a species head combines trait embeddings with context. Stop-gradient operations at the boundary between stages ensure that species and trait losses never back-propagate into lower-level representations. Experi

What carries the argument

The load-bearing mechanism is the unidirectional information-flow chain enforced by stop-gradient (gradient-detach) operations, instantiated from the factorization P(y,t,M|X)=P(y|t,X)P(t|M,X)P(M|X). Segmentation evidence, summarized as max-pooled prototype responses, is concatenated into the trait head with gradients blocked; trait embeddings are then detached and fed to the species head. Around this chain sit two supporting devices: a prototype-based segmentation head whose learnable prototypes are regularized toward orthogonality, and homoscedastic uncertainty weighting that adaptively down-weights the species loss because its gradients are treated as high-variance noisy signals.

Load-bearing premise

The central premise is that species identity is mediated by the annotated morphological traits — that once traits are known, segmentation supplies no extra species information — and that those trait labels are reliable enough to carry the bridge.

What would settle it

Shuffle or corrupt the trait labels while keeping segmentation intact: if ultra-rare accuracy still climbs roughly 13 points over the baseline, the trait bridge contributes nothing and the gain must come from some other component. Alternatively, train FISHER on a class-balanced subset: if the advantage over parallel MTL disappears, the claim that gradient conflict is the bottleneck is not general.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, the trait bridge lets ultra-rare species be recognized from shared morphological cues rather than memorized global patterns, so long-tail accuracy should improve on any fine-grained dataset with intermediate attributes.
  • The gradient-alignment analysis predicts that parallel MTL baselines will show persistent negative cosine similarity between classification and segmentation gradients; FISHER keeps this positive, so the mechanism is measurable during training.
  • Because the gains come from structure rather than capacity, the method should transfer to smaller backbones without losing the long-tail benefit.
  • The Leave-Out-Species result implies the hierarchy learns species-agnostic trait representations that generalize to unseen taxa, which matters for open-set biodiversity monitoring.

Where Pith is reading between the lines

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

  • The four binary traits (adipose, pelvic, barbel, dorsal) may be too coarse to capture all species-discriminative anatomy; the same architecture with a richer trait vocabulary could push the rare-species gains further — a natural extension the paper does not test.
  • The decoupling principle is not fish-specific: any domain with a part→attribute→category hierarchy (plant identification, medical pathology, bird morphology) could adopt the same stop-gradient chain.
  • One testable extension is to vary the strength of gradient coupling continuously instead of hard stop-gradients; the paper's ablation suggests even partial connectivity hurts OOD trait mAP by roughly ten points, implying an optimal-coupling regime worth mapping.
  • The uncertainty-weighting analysis reveals the species loss is down-weighted to near zero; an inverse inference is that noisy long-tail labels may be actively harmful to shared features, so a label-cleaning or robust-loss variant might compound the gains.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes FISHER, a hierarchical multi-task architecture for fine-grained aquatic species recognition on the Fish-Vista benchmark. The model factorizes the prediction problem as segmentation → trait prediction → species classification, and enforces a unidirectional flow by applying gradient detachment at the task interfaces, so that higher-level losses do not backpropagate into lower-level representations. The segmentation head uses learnable prototypes with an orthogonality regularizer, and the three task losses are combined with homoscedastic uncertainty weighting. Experiments report improved species classification, particularly for ultra-rare classes (63.8% vs. 50.4% for Swin-B-22k), high OOD trait mAP (97.7% on the Leave-Out-Species split), and ablation results indicating that gradient conflicts degrade segmentation and trait generalization.

Significance. The paper addresses an important and timely problem—long-tailed fine-grained recognition in biodiversity monitoring—and its core idea is well motivated. The ablation pattern in Table VIII and the comparison with gradient-based MTL methods in Table IX provide useful evidence that negative transfer exists in this setting and that structural gradient decoupling is a plausible remedy. If the decoupling-specific gains are confirmed with repeated runs, the work would be a valuable contribution to hierarchical multi-task learning and fine-grained recognition. The manuscript is transparent about its components and reports a broad set of baselines.

major comments (4)
  1. [V-B/V-C, Tables III, VIII, IX] The headline claim '+13.4% ultra-rare accuracy' compares FISHER to a species-only Swin-B classifier. Fig. 4 shows that the uncertainty weight assigned to the species loss is below 1% while the trait loss receives ~90% of the total weight, so most of the gain over Swin-B can be explained by auxiliary segmentation/trait supervision rather than by gradient decoupling. The decoupling-specific comparisons are FISHER vs. Fully Connected in Table VIII (63.8 vs. 61.4) and vs. MTL+UW in Table IX (63.8 vs. 60.4). These margins are 2.4 and 3.4 points, and all results are single runs with no error bars, seeds, or significance tests. This is not sufficient to establish that the novel detachment mechanism causes the gain; report multi-seed means with standard deviations for at least Tables VIII and IX and for Fig. 5.
  2. [IV-B/IV-C, Eq. (12)] The species head receives bc=[g;p_detach;h_detach_t], so the global descriptor g and the segmentation evidence p are available to the classifier in addition to the trait embedding h_t. The factorization in Eq. (2) therefore does not enforce a trait bottleneck, and the claim that trait representations are the critical bridge to ultra-rare species is not directly tested. Please add an experiment in which the species head uses only the detached trait embedding (or in which g is ablated) to show that the hierarchy is load-bearing; otherwise the No-Trait ablation in Table VIII may be attributable to extra feature capacity rather than to the trait bridge.
  3. [IV-D Eq. (22), V-C Table VIII, Fig. 4] The uncertainty-weighting description is internally inconsistent. Eq. (22) gives a weight exp(-eta_m)=1/sigma_m^2, while Fig. 4 reports weights as 1/(2 sigma_m^2). More importantly, Table VIII shows that adding uncertainty weighting alone to the detached base reduces Species F1 from 58.1 to 55.7 and ultra-rare accuracy from 61.5 to 59.6, yet the text states that removing UW makes the species loss dominate and decreases ultra-rare accuracy from 63.8 to 61.5. Please correct the formula and characterize the actual task-loss/gradient magnitudes so the claimed interaction between UW and orthogonality is supported.
  4. [III-A Eq. (2), V-I Fig. 10] The hierarchical factorization assumes that the four binary traits are sufficient to connect segmentation and species classification. The paper does not analyze trait annotation quality or trait informativeness; the high Adipose-Barbel prototype correlation (0.67) in Fig. 10 suggests these traits may be difficult to separate. Since the biological-hierarchy motivation depends on the trait bridge, please report per-trait reliability/agreement or an experiment that uses predicted traits as the sole input to the species head, otherwise the sufficiency assumption is unverified.
minor comments (5)
  1. [Throughout, esp. V-C and Fig. 3] The manuscript contains multiple uninterpretable '/uni0000...' escape sequences (e.g., after 'C. Ablation Study' and in the Fig. 3 caption) that render parts of the text unreadable. The file must be re-typeset cleanly before any further review.
  2. [Abstract and Tables V-VI] The abstract says 'unseen trait identification,' but the experiments evaluate traits on Leave-Out-Species (unseen species), not on unseen traits. Please reword to avoid confusion.
  3. [Table VIII] For the 'Fully Connected' and 'Standard MTL' baselines, specify whether they use the same prototype segmentation head, orthogonality loss, and uncertainty weighting as FISHER. Without this, the ablation does not cleanly isolate gradient detachment.
  4. [V-F, Fig. 5] The curves labeled 'Baseline (Connected)', 'Baseline (Parallel)', and 'Orthogonal (Target)' are not all defined in the text. Define each variant, state the smoothing window, and report the number of runs used for the alignment curves.
  5. [V-A, Eq. (31)] Species F1 is reported as a macro-F1, but no formula or definition is given for species macro-F1. Please add the definition or a reference.

Circularity Check

0 steps flagged

No significant circularity: FISHER's performance claims are tested against external baselines and held-out frequency groups; the hierarchy equation is an assumption, not a fitted prediction.

full rationale

The paper's central claim is an architectural/empirical one: gradient-decoupled hierarchical multi-task learning improves fine-grained aquatic species recognition. The factorization in Eq. (1)-(2) is a modeling assumption rather than a derived prediction, and the architecture is evaluated against external baselines (Swin-B-22k, Q2L, MaxViT, etc.) on the Fish-Vista benchmark, including held-out unseen species and manually annotated data. The ultra-rare species accuracy gap (63.8% vs. 50.4%) is not equivalent to any fitted parameter or to the hierarchical factorization itself; it is an externally measured result. Ablations include connected/parallel baselines that do not use detachment, and the gradient-alignment analysis (Fig. 5) is a diagnostic observation, not a target that was fit. No load-bearing self-citations appear; citations to Fish-Vista, Swin, uncertainty weighting, PCGrad, and SimSiam are external prior work, not the authors' own unverified results. The single-run nature of the experiments and the absence of significance tests are legitimate correctness/robustness concerns but do not constitute circularity under the stated criteria.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central structural assumptions are the factorization of the task hierarchy and the sufficiency of the four traits. The free parameters are mostly standard hyperparameters, but λ_ortho is explicitly tuned on the validation metric, and the uncertainty weights are learned and drive the task balance. No new physical entities are introduced.

free parameters (4)
  • Prototype temperature τ = 0.1
    Set by hand (Table II); controls sharpness of cosine similarity in Eq. (4). No sensitivity analysis.
  • Orthogonality weight λ_ortho = 0.5
    Tuned via sweep over [0.0, 1.0] (Fig. 3); the paper selects 0.5 as the peak of ultra-rare accuracy and mIoU.
  • Label smoothing ε = 0.1
    Set by hand (Table II); standard practice, but no sensitivity analysis.
  • Uncertainty weights η_m = Learned; final relative weights shown in Fig. 4
    Learnable parameters in Eq. (22); the paper reports that species weight becomes <1%, trait weight ~90%. This is a fitted parameter affecting task balance.
axioms (4)
  • domain assumption Species classification can be factorized through traits: P(y|t,M,X) = P(y|t,X)
    The model assumes segmentation evidence matters only through trait predictions (Eq. 1-2). If segmentation carries unique discriminative information not captured by the four traits, this information is thrown away.
  • domain assumption The four binary traits (adipose, pelvic, barbel, dorsal) are sufficient to bridge segmentation and species classification
    The trait head uses only four traits. If two species differ in a trait not in this set, the bridge cannot distinguish them.
  • ad hoc to paper Gradient interference between dense and high-level tasks is the primary cause of negative transfer in this setting
    The paper asserts this as the core hypothesis and supports it with ablations and gradient cosine analysis, but it is not proven as a universal mechanism.
  • domain assumption Swin-B pre-trained on ImageNet-22k provides a suitable shared backbone for aquatic recognition
    The backbone is not trained from scratch; the paper assumes ImageNet-22k features transfer to fish morphology.
invented entities (1)
  • Learnable prototype matrix P in the segmentation head independent evidence
    purpose: Represents each anatomical class as a vector in feature space; used for cosine-similarity segmentation logits.
    The prototypes are learned from data and their quality is evaluated via segmentation mIoU and the correlation matrix in Fig. 10. They are not a new physical entity, just parameters.

pith-pipeline@v1.3.0-alltime-deepseek · 22635 in / 8046 out tokens · 70229 ms · 2026-08-02T07:53:48.421801+00:00 · methodology

0 comments
read the original abstract

Fine-grained recognition of aquatic species is challenging due to subtle morphological differences and long-tailed distributions, where ultra-rare species are underrepresented. A natural solution is to jointly model segmentation, morphological traits, and species classification within a multi-task learning (MTL) framework. However, existing MTL methods suffer from negative transfer caused by gradient conflicts between low-level dense tasks and high-level classification objectives, degrading fine-grained representations. To address this limitation, we identify gradient interference across hierarchical tasks as a fundamental bottleneck and propose FISHER, a gradient-decoupled hierarchical multi-task learning framework. FISHER aligns optimization with the biological hierarchy of aquatic species by enforcing a unidirectional information flow from segmentation to trait prediction and finally to species classification, while explicitly decoupling gradients across task boundaries. This design prevents high-level objectives from corrupting low-level morphological representations, effectively mitigating negative transfer while preserving the benefits of shared supervision. Furthermore, we introduce a prototype-based segmentation head with orthogonality regularization to encourage disentangled anatomical representations, and employ homoscedastic uncertainty weighting to dynamically balance task contributions during training. Our analysis shows that robust trait representations serve as a critical bridge for transferring knowledge to ultra-rare species. Extensive experiments on the Fish-Vista benchmark demonstrate that FISHER achieves 97.7% mAP for unseen trait identification and improves ultra-rare species classification accuracy by 13.4% over strong baselines, highlighting the effectiveness of gradient-decoupled hierarchical learning for long-tailed biodiversity recognition.

Figures

Figures reproduced from arXiv: 2607.20523 by Ba Hung Ngo, Cuong D. Do, Mai Phuong Tran, Phuc H. Nguyen, Van-Dinh Nguyen.

Figure 1
Figure 1. Figure 1: Comparison of parallel and hierarchical multi-task learning. Parallel [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the FISHER architecture. Features are extracted by a shared backbone and refined hierarchically from segmentation to trait prediction and species classification. Red dashed lines denote gradient decoupling (stop-gradient) used to mitigate negative transfer. frameworks often integrate multiple objectives; for example, YOLO-FD [19] combines detection and segmentation for fish disease analysis, wh… view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis of orthogonality weight ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Gradient Alignment in Shared Backbone. Cosine similarity between [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Grad-CAM visualization results on a notropis rubellus sample. From left to right: Original Image, FISHER (Ours), and ablation baselines: ClsOnly (classification only), NoSeg (without segmentation branch), NoTraits (without attribute-aware prototypes), and Reverse (reversed task hierarchy). Original Image Ground Truth Prediction Overlay Head Eye Dorsal Pectoral Pelvic Anal Caudal Adipose Barbel [PITH_FULL_… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of segmentation masks generated by FISHER. This behavior is further supported by the ablation study in Table VIII. Removing uncertainty weighting (i.e., using fixed equal weights) causes the species loss to dominate opti￾mization, leading to reduced ultra-rare accuracy (63.8% → 61.5%) and slightly degraded segmentation quality (Barbel IoU: 32.9% → 32.6%). These results highlight that dynam￾ic… view at source ↗
Figure 8
Figure 8. Figure 8: Visualization results of trait identification extracted by the proposed FISHER. (a) ortho =0.0 Compactness: 70.81 | Separation: 68.79 Head Eye Dorsal Pectoral Pelvic Anal Caudal Adipose Barbel (b) ortho =0.5 Compactness: 51.95 | Separation: 52.51 Head Eye Dorsal Pectoral Pelvic Anal Caudal Adipose Barbel (c) ortho =1.0 Compactness: 63.36 | Separation: 58.96 Head Eye Dorsal Pectoral Pelvic Anal Caudal Adipo… view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE visualization of the semantic features to illustrate the effectiveness of the orthogonality constraint for [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Ultra-rare species accuracy versus model complexity. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 1 linked inside Pith

  1. [1]

    Fish-vista: A multi-purpose dataset for understanding & identification of traits from images,

    K. S. Mehrab, M. Maruf, A. Dawet al., “Fish-vista: A multi-purpose dataset for understanding & identification of traits from images,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2025, pp. 24 275–24 285. NGUYENet al.: GRADIENT-DECOUPLED HIERARCHICAL MULTI-TASK LEARNING FOR FINE-GRAINED AQUATIC SPECIES RECOGNITION 13

    2025

  2. [2]

    Autofish: Dataset and benchmark for fine-grained analysis of fish,

    S. H. Bengtson, D. Lehotsk ´y, V . Ismiroglou, N. Madsen, T. B. Moeslund, and M. Pedersen, “Autofish: Dataset and benchmark for fine-grained analysis of fish,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), February 2025, pp. 1598–1607

    2025

  3. [3]

    M4FT: Mamba, migratory, mobile and multiple fish tracking,

    W. Li, Y . Liu, W. Wang, and Z. Li, “M4FT: Mamba, migratory, mobile and multiple fish tracking,”IEEE Trans. Circuit Syst. Video Technol., pp. 1–1, 2026

    2026

  4. [4]

    Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction,

    F. F. Khanet al., “Fishnet: A large-scale dataset and benchmark for fish recognition, detection, and functional trait prediction,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), October 2023, pp. 20 496– 20 506

    2023

  5. [5]

    A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis,

    A. Saleh, I. H. Laradji, D. A. Konovalov, M. Bradley, D. Vazquez, and M. Sheaves, “A realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis,”Scientific Reports, vol. 10, no. 1, p. 14671, 2020

    2020

  6. [6]

    The deepfish computer vision dataset for fish instance segmentation, classification, and size estimation,

    N. Garcia-d’Ursoet al., “The deepfish computer vision dataset for fish instance segmentation, classification, and size estimation,”Scientific Data, vol. 9, no. 1, p. 287, 2022

    2022

  7. [7]

    Automatic estuarine fish species classification system based on deep learning techniques,

    H. Tejaswini, M. M. Manohara Pai, and R. M. Pai, “Automatic estuarine fish species classification system based on deep learning techniques,” IEEE Access, vol. 12, pp. 140 412–140 438, 2024

    2024

  8. [8]

    Fish tracking and segmentation from stereo videos on the wild sea surface for electronic monitoring of rail fishing,

    T.-W. Huang, J.-N. Hwang, S. Romain, and F. Wallace, “Fish tracking and segmentation from stereo videos on the wild sea surface for electronic monitoring of rail fishing,”IEEE Trans. Circuit Syst. Video Technol, vol. 29, no. 10, pp. 3146–3158, 2019

    2019

  9. [9]

    Fish tracking and continual behavioral pattern clustering using novel sillago sihama vid (ssvid),

    S. Shreesha, M. M. M. Pai, U. Verma, and R. M. Pai, “Fish tracking and continual behavioral pattern clustering using novel sillago sihama vid (ssvid),”IEEE Access, vol. 11, pp. 29 400–29 416, 2023

    2023

  10. [10]

    Fishnet: Fish visual recognition with one stage multi-task learning,

    Z. Chen, L. Cao, Q. Wang, and Y . Cai, “Fishnet: Fish visual recognition with one stage multi-task learning,”IET Image Processing, vol. 16, no. 12, pp. 3237–3246, 2022

    2022

  11. [11]

    Mtaffb: a multi-task active learning framework for analyzing fish school feeding behavior,

    T. Zhou, X. Peng, and Y . Zhang, “Mtaffb: a multi-task active learning framework for analyzing fish school feeding behavior,”Smart Agricul- tural Technology, vol. 12, p. 101313, 2025

    2025

  12. [12]

    Exploring simple siamese representation learning,

    X. Chen and K. He, “Exploring simple siamese representation learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021

    2021

  13. [13]

    Interpretable image classi- fication via non-parametric part prototype learning,

    Z. Zhu, L. Fan, M. Pagnucco, and Y . Song, “Interpretable image classi- fication via non-parametric part prototype learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2025, pp. 9762– 9771

    2025

  14. [14]

    Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,

    R. Cipolla, Y . Gal, and A. Kendall, “Multi-task learning using uncer- tainty to weigh losses for scene geometry and semantics,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 7482–7491

    2018

  15. [15]

    Wildfish++: A comprehensive fish benchmark for multimedia research,

    P. Zhuang, Y . Wang, and Y . Qiao, “Wildfish++: A comprehensive fish benchmark for multimedia research,”IEEE Trans. Multi., vol. 23, p. 3603–3617, 2021

    2021

  16. [16]

    Fine-grained fish classification from small to large datasets with vision transformers,

    R. J. M. Veiga and J. M. F. Rodrigues, “Fine-grained fish classification from small to large datasets with vision transformers,”IEEE Access, vol. 12, pp. 113 642–113 660, 2024

    2024

  17. [17]

    A lightweight fine-grained pelagic fish recognition algorithm based on object detec- tion,

    W. Ren, D. Shi, Y . Chen, L. Song, Q. Hu, and M. Wang, “A lightweight fine-grained pelagic fish recognition algorithm based on object detec- tion,”Aquaculture International, vol. 33, no. 2, Jan 2025

    2025

  18. [18]

    BioCLIP: A vision foundation model for the tree of life,

    S. Stevenset al., “BioCLIP: A vision foundation model for the tree of life,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024, pp. 19 412–19 424

    2024

  19. [19]

    Yolo-fd: An accurate fish disease detection method based on multi-task learning,

    X. Li, S. Zhao, C. Chen, H. Cui, D. Li, and R. Zhao, “Yolo-fd: An accurate fish disease detection method based on multi-task learning,” Expert Systems with Applications, vol. 258, p. 125085, Dec 2024

    2024

  20. [20]

    Deepseavision: Enhanced detection and classification of underwater species,

    G. Joshita Reddy, K. Reddy, S. Ruthvik Athota, S. J. Narayanan, B. Perumal, and G. Kumar Nayak, “Deepseavision: Enhanced detection and classification of underwater species,”IEEE Access, vol. 13, pp. 173 347–173 367, 2025

    2025

  21. [21]

    Optimizing fish classification with a hybrid sae-svm model: Perfor- mance and evaluation on fish-gres and fish4-knowledge datasets,

    S. Senthil Pandi, R. Suguna Devi, A. Thanam, and S. M. Keerthana, “Optimizing fish classification with a hybrid sae-svm model: Perfor- mance and evaluation on fish-gres and fish4-knowledge datasets,” in Proc. Int. Conf. Recent Adv. Sci. Eng. Technol. (ICRASET), 2024, pp. 1–6

    2024

  22. [22]

    Which tasks should be learned together in multi-task learning?

    T. Standley, A. Zamir, D. Chen, L. Guibas, S. Savarese, and W. Saxe, “Which tasks should be learned together in multi-task learning?” inProc. Int. Conf. Mach. Learn. (ICML), 2020, pp. 9120–9132

    2020

  23. [23]

    Model- protected multi-task learning,

    J. Liang, Z. Liu, J. Zhou, X. Jiang, C. Zhang, and F. Wang, “Model- protected multi-task learning,”IEEE Trans. Patt. Anal. and Mach. Intell., vol. 44, no. 2, p. 1002–1019, Feb 2022

    2022

  24. [24]

    Decoupling representation learning and classifier for long- tailed adversarial training,

    H. Xiong, D. Man, J. Lv, C. Xu, F. Zeng, Y . Shi, M. Lai, and W. Yang, “Decoupling representation learning and classifier for long- tailed adversarial training,”Pattern Recognition, vol. 172, p. 112607, Apr 2026

    2026

  25. [25]

    Gradient surgery for multi-task learning,

    T. Yu, S. Kumar, A. Gupta, S. Levine, K. Hausman, and C. Finn, “Gradient surgery for multi-task learning,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2020

    2020

  26. [26]

    Individual fish recognition method with coarse and fine-grained feature linkage learning for precision aquaculture,

    J. Yin, J. Wu, C. Gao, H. Yu, L. Liu, Z. Jiang, and S. Guo, “Individual fish recognition method with coarse and fine-grained feature linkage learning for precision aquaculture,”Aquaculture Research, vol. 2023, p. 1–14, 2023

    2023

  27. [27]

    Hierarchical feature attention learning network for detecting object and discriminative parts in fine-grained visual classification,

    A. Yeong Han, K. Moo Yi, K. Tae Kim, and J. Young Choi, “Hierarchical feature attention learning network for detecting object and discriminative parts in fine-grained visual classification,”IEEE Access, vol. 13, pp. 19 533–19 544, 2025

    2025

  28. [28]

    Concept bottleneck models,

    P. W. Kohet al., “Concept bottleneck models,” inProc. Int. Conf. Mach. Learn. (ICML), 2020, pp. 5338–5348

    2020

  29. [29]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” inProc. Int. Conf. Learn. Represent. (ICLR), 2015

    2015

  30. [30]

    Aggregated residual transformations for deep neural networks,

    S. Xie, R. Girshick, P. Doll ´ar, Z. Tu, and K. He, “Aggregated residual transformations for deep neural networks,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017, pp. 1492–1500

    2017

  31. [31]

    Designing network design spaces,

    I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, and P. Doll ´ar, “Designing network design spaces,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2020, pp. 10 428–10 436

    2020

  32. [32]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. Int. Conf. Learn. Represent. (ICLR), 2021

    2021

  33. [33]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liuet al., “Swin transformer: Hierarchical vision transformer using shifted windows,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021, pp. 10 012–10 022

    2021

  34. [34]

    Maxvit: Multi-axis vision transformer,

    Z. Tu, H. Talebi, H. Zhang, F. Yang, P. Milanfar, A. Bovik, and Y . Li, “Maxvit: Multi-axis vision transformer,” inProc. Eur. Conf. Comput. Vis. (ECCV). Springer, 2022, pp. 459–479

    2022

  35. [35]

    Learning transferable visual models from natural language supervision,

    A. Radfordet al., “Learning transferable visual models from natural language supervision,” inProc. Int. Conf. Mach. Learn. (ICML). PMLR, 2021, pp. 8748–8763

    2021

  36. [36]

    DINOv2: Learning robust visual features without supervision,

    M. Oquabet al., “DINOv2: Learning robust visual features without supervision,”Transactions on Machine Learning Research, 2024

    2024

  37. [37]

    A simple interpretable transformer for fine-grained image classification and analysis,

    D. Paulet al., “A simple interpretable transformer for fine-grained image classification and analysis,” inProc. Int. Conf. Learn. Represent. (ICLR), 2024

    2024

  38. [38]

    Transfg: A transformer architecture for fine-grained recognition,

    J. He, J.-N. Chen, S. Liu, A. Kortylewski, C. Yang, Y . Bai, and C. Wang, “Transfg: A transformer architecture for fine-grained recognition,” in Proc. AAAI Conf. Artif. Intell. (AAAI), vol. 36, 06 2022, pp. 852–860

    2022

  39. [39]

    Pyramid scene parsing network,

    H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing network,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2017

    2017

  40. [40]

    Encoder- decoder with atrous separable convolution for semantic image segmen- tation,

    L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder- decoder with atrous separable convolution for semantic image segmen- tation,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2018

    2018

  41. [41]

    Per-pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,” inAdv. Neural Inf. Process. Syst. (NeurIPS), 2021

    2021

  42. [42]

    Yolov8: A novel object detection algorithm with enhanced performance and robustness,

    R. Varghese and M. Sambath, “Yolov8: A novel object detection algorithm with enhanced performance and robustness,” inProc. Int. Conf. Adv. Data Eng. Intell. Comput. Syst. (ADICS). IEEE, 2024, pp. 1–6

    2024

  43. [43]

    Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models,

    M. Deitkeet al., “Molmo and pixmo: Open weights and open data for state-of-the-art vision-language models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), June 2025, pp. 91–104

    2025

  44. [44]

    Query2label: A simple transformer way to multi-label classification,

    S. Liu, L. Zhang, X. Yang, H. Su, and J. Zhu, “Query2label: A simple transformer way to multi-label classification,”arXiv preprint arXiv:2107.10834, 2021

    Pith/arXiv arXiv 2021

  45. [45]

    GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks,

    Z. Chen, V . Badrinarayanan, C.-Y . Lee, and A. Rabinovich, “GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks,” inProc. Int. Conf. Mach. Learn. (ICML), 2018

    2018

  46. [46]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inProc. IEEE Int. Confe. Comp. Vis. (ICCV), Oct 2017

    2017

  47. [47]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”J. Mach. Lear. Res., vol. 9, no. 11, p. 2579–2605, 2008

    2008