Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Object-Centric Action-Enhanced Representations for Robot Visuo-Motor Policy Learning

T0 review · 4 major / 5 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 An object-centric encoder fine-tuned on human action videos produces visual representations that improve robot visuo-motor policies on the TOTO simulated pouring task.

desk verdict Plausible fine-tuning signal from human action videos, but the object-centric attribution is confounded by a missing control and tuning on the benchmark. read the letter →

arxiv 2505.20962 v1 pith:OEYEOAO6 submitted 2025-05-27 cs.RO cs.CV

Nikos Giannakakis , Argyris Manetas , Panagiotis P. Filntisis , Petros Maragos , George Retsinas This is my paper
classification cs.ROcs.CV
keywords object-centricrepresentationlearningslotattentionrobotmanipulationbehaviorcloningofflinereinforcementvisualTOTObenchmarkaffordances
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

This paper claims that a robot's visual state representation is more useful for learning manipulation policies if it is built from object-level slots rather than one global image vector, and if the slot encoder is first fine-tuned on videos of humans performing actions. The authors adapt SOLV, a video object-centric model, by keeping its slot attention module and removing its temporal binding step, then fine-tune it on a subset of the Something-Something action dataset. On the TOTO simulated pouring task, the resulting action-enhanced encoder gives the best success rate and mean reward among the compared encoders under both behavior cloning (0.61, 25.25) and offline reinforcement learning via implicit Q-learning (0.57, 24.41). If the claim holds, large and inexpensive human action video collections could substitute for robot-specific or annotated data when pretraining visual encoders for manipulation.

What carries the argument

The central machinery is the adapted SOLV encoder with action enhancement, A-SOLV (AE): a Slot Attention module with shared initialization operating on frozen DINOv2 image embeddings, followed by a Slot Merger that uses agglomerative clustering to output exactly four object-specific vectors. Each slot yields a 128-dimensional "what" vector for object features, while its attention mask, resized to 10 by 10 and passed through scaled softmax, is flattened into a 100-dimensional "where" vector for location. The what and where vectors are concatenated into a 912-dimensional scene representation that forms the visual part of the policy state. A deliberate design choice is the removal of SOLV's temporal binding component; the authors argue that the slot vectors already retain nontrivial temporal information from pretraining, so the component is unnecessary. The component ablations, varying the number of slots and including or excluding the where vectors, are used to justify these design choices.

What would settle it

Train A-SOLV (AE) with SOLV's temporal binding component retained, and in a separate arm replace the Something-Something fine-tuning with fine-tuning on a static-image dataset of equal size; if either variant matches the reported 0.61 success rate, the paper's explanation in terms of action-aware temporal slot vectors is not supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that coupling semantic segmentation and representation generation in one object-centric encoder, and bootstrapping that encoder from a video object-centric model fine-tuned on human action videos, yields visual state representations that transfer to robot visuo-motor policies better than representations from standard image encoders. The encoder runs slot attention on frozen DINOv2 image embeddings, merges slots into four object-specific "what" vectors of 128 dimensions, and appends a 100-dimensional "where" vector per slot derived from its attention mask, producing a 912-dimensional scene representation. In the TOTO simulated pouring task, behavior-cloning agents using this encoder reach a 0.61 success rate and 25.25 mean reward, and implicit Q-learning agents reach 0.57 and 24.41, the best among the compared encoders in both settings. The paper attributes the gain to temporally infused object features and to fine-tuning on human action videos, which aligns the representation with the affordances and dynamics relevant to robot manipulation.

Load-bearing premise

The load-bearing premise is that the slot vectors produced by slot attention alone retain enough temporal and action information after the temporal binding component is removed, an assumption the paper argues for rather than demonstrates; if it fails, the measured improvement could come from fine-tuning alone rather than from action-aware object representations.

Editorial extensions

If this is right

  • Robot visuo-motor policies can be bootstrapped from visual encoders pretrained on large, annotation-free human action video collections, reducing reliance on robot-specific or hand-labeled data.
  • An encoder that jointly performs segmentation and object-feature extraction can be used directly as policy input at a representation size comparable to or smaller than standard image encoders.
  • Dropping the explicit spatial "where" vectors degrades performance, so manipulation policies benefit from encoding object location as well as appearance.
  • The gap between the action-enhanced and generic-video variants indicates that the choice of fine-tuning data matters, not just the object-centric architecture.

Reading between the lines

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

  • Beyond the paper, the slot vectors could be probed on held-out action clips for action predictability; such a probe would test whether temporal information survives the removal of the temporal binding component, a question the paper leaves open.
  • Beyond the paper, the fixed four-slot merger should be re-examined on scenes with more than four objects or heavy occlusion, where under-segmentation would likely hurt policy performance.
  • Beyond the paper, if human-action alignment is the active ingredient, fine-tuning on egocentric manipulation videos should yield larger gains than third-person Something-Something clips, and the improvement should scale with dataset size and diversity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an object-centric visual encoder, A-SOLV, built on the SOLV slot-attention model. It uses a frozen DINOv2 backbone, slot attention, and slot merging to produce four slots, then concatenates 128-dimensional 'what' features with 100-dimensional 'where' location vectors derived from attention masks. The encoder is fine-tuned for 100 epochs on a subset of the Something-Something action video dataset. The authors evaluate the resulting 912-dimensional representation on the TOTO simulated pouring task, training Behavior Cloning (BC) and Implicit Q-Learning (IQL) policies and comparing against frozen BYOL, CLIP, DINOv2, MoCo, ResNet50, and an A-SOLV version without action fine-tuning. They report the highest success rate and mean reward for A-SOLV(AE) in both settings (Table I: 0.61 success, 25.25 mean reward; Table II: 0.57 success, 24.41 mean reward) and conclude that object-centric action-enhanced representations improve robot visuo-motor policy learning.

Significance. If substantiated, the work would offer a practical recipe for leveraging out-of-domain human action videos to build compact object-centric state representations for robot manipulation without robot-specific pretraining data. The representation size (912) is smaller than ResNet50 and MoCo, and the use of a frozen DINOv2 backbone is computationally accessible. The authors should be credited for using five seeds and 100 evaluation trajectories in the BC experiments, and for directly comparing the action-fine-tuned and non-fine-tuned versions of their encoder. However, the current evidence does not isolate the claimed mechanism or establish a statistically robust advantage over frozen baselines; the main gap is the absence of a fine-tuned, non-object-centric control.

major comments (4)
  1. [§III-A, Table I] The central attribution claim is confounded by fine-tuning. A-SOLV(AE) is fine-tuned for 100 epochs on a Something-Something subset, while every comparison encoder (BYOL, CLIP, DINOv2, MoCo, ResNet50) is used as a fixed pretrained extractor. The comparison between A-SOLV* and A-SOLV(AE) isolates the effect of action-video fine-tuning within the object-centric family, but the comparison against frozen baselines cannot separate the effect of object-centric slot structure from the effect of additional task-aligned fine-tuning. Please add at least one non-object-centric baseline (e.g., DINOv2 or ResNet50) fine-tuned on the exact same Something-Something subset for the same number of epochs and evaluated under the same BC/IQL protocol. Without this control, the reported success-rate gain may be due to fine-tuning alone.
  2. [Tables I and II] No statistical significance testing is reported. In Table I, A-SOLV(AE) achieves 0.61 ± 0.04 success versus ResNet50's 0.56 ± 0.11 and DINOv2's 0.55 ± 0.04; with five seeds these distributions overlap substantially, so the statement in Section III-C that the method 'consistently achieves the highest reward and success rate' is not supported without a formal test (e.g., bootstrap confidence intervals, a two-sample test across seeds, or effect sizes). Table II reports only point estimates and no error bars or number of seeds, so the IQL comparison (0.57 vs 0.46 for ResNet50) cannot be assessed.
  3. [§III-E] The number of slots is selected on the same benchmark used for evaluation. Section III-E states that the slot count was chosen after 'a re-evaluation of the imitation learning experiments' on the TOTO pouring task (Figure 5). This is test-set tuning and can inflate the reported success rate; the same concern applies to the scaled-softmax adjustment in Section II and the 100-epoch fine-tuning schedule. Please use a separate validation split or nested evaluation, and report performance for all slot counts tried (4, 6, 8) with the selection rule stated in advance.
  4. [§II] The decision to remove SOLV's Temporal Binding Component is justified only by an argument, not by evidence: 'we argue that the slot vectors produced by the slot attention module already capture non-trivial temporal information' (Section II). Since the paper's central claim is that the representations are action-enhanced, this architectural choice needs an ablation that includes the temporal binding component, or otherwise measures whether the resulting slots actually preserve temporal and affordance information. Without such evidence, the mechanism behind the improvement over A-SOLV* remains unclear.
minor comments (5)
  1. [§III-B] The paper states that TOTO includes pouring and scooping tasks, but all experiments are on pouring only; the title and abstract claim broader robot visuo-motor policy learning. Please either add the scooping task or qualify the claims to the pouring task.
  2. [§III-A] The Something-Something subset is underspecified (size, class distribution, balance, train/validation split), which limits reproducibility of the fine-tuning step.
  3. [§III-F] Figure 6 compares 'what' only with 'what'+'where' representations but reports no numerical values or error bars; please include the underlying success rates and rewards in the text or a table.
  4. [Section II] There is a typo: 'temporary binding step' should be 'temporal binding step.'
  5. [Table II] Table II should report the number of seeds and variance or confidence intervals, as Table I does, to allow the reader to judge the reliability of the IQL comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the performance claims are empirical outcomes from an external benchmark, not constructed from the method's definitions or from self-citations.

full rationale

No circular derivation chain is present. The paper's method, A-SOLV (AE), is an encoder built from SOLV's slot attention mechanism and fine-tuned on a Something-Something subset; its downstream effect on BC and IQL policies is measured by simulation rollouts on the external TOTO benchmark, where rewards and trajectories are fixed independently of the authors' definitions. The reported success rates and mean rewards are empirical outcomes, not quantities recovered by construction from the encoder's weights or from any fitted parameter renamed as a prediction. The architecture choices (four slots, where-vector concatenation, scaled softmax, and the ablation-based slot count) are tuned through experiments, including a re-evaluation on the same TOTO task in Section III-E; this is a test-set selection concern rather than circularity, because the final results are not logically forced to equal the selection criterion. The paper's use of SOLV, DINOv2, and TOTO as external references is independent support rather than a load-bearing self-citation. The argument that slot vectors retain temporal information after removing SOLV's temporal binding component (Section II) is an unproven modeling assumption, not a circular step, since it neither defines nor derives the reported policy improvements. The missing control of a fine-tuned non-object-centric baseline is an experimental attribution issue, not a circularity issue. The central claim therefore rests on externally measured results and does not reduce to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities; it only combines existing components into a new encoder. The free parameters are the design choices tuned on the benchmark, and the axioms are domain assumptions about transfer and the suitability of the pretrained components.

free parameters (4)
  • Number of final slots = 4
    Chosen after experiments comparing 4, 6, and 8 slots on the TOTO benchmark (Section III-E); affects segmentation granularity and representation size.
  • Where vector grid resolution = 10 x 10
    Attention masks are resized to 10x10 before flattening into spatial vectors (Section II); this resolution is hand-chosen and not ablated.
  • Scaled softmax scaling = unspecified
    A scaled softmax is applied to resized attention masks and described as improving performance (Section II), but the scaling factor is not reported.
  • Fine-tuning epochs on Something-Something = 100
    Fine-tuning duration on the action video subset (Section III-A); no sweep or analysis is provided.
assumptions (5)
  • domain assumption Slot attention binds slots to semantically meaningful object regions without supervision
    Relies on results from [17], [18]; the quality of object segmentation is not independently verified for the robot domain.
  • domain assumption DINOv2 features are a suitable frozen visual backbone for slot attention
    The frozen DINOv2 encoder provides input embeddings (Section II); no comparison with other backbones is made.
  • domain assumption Fine-tuning on Something-Something human action videos transfers to robotic manipulation tasks
    The paper assumes action-object associations from human videos align with the TOTO pouring task (Section III-A); this is the core transfer premise.
  • ad hoc to paper Removing SOLV's temporal binding does not lose temporally useful information
    Section II states the slot vectors 'already capture non-trivial temporal information' without supporting experiments; this is a specific assumption of the proposed modification.
  • domain assumption Object-centric representations are more effective than global image features for robot policy learning
    Motivated by psychological theories (Section I) and prior works [15], [16]; used to justify the architecture choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Object-Centric Action-Enhanced Representations for Robot Visuo-Motor Policy Learning." pith.science (2026). https://pith.science/paper/OEYEOAO6

@misc{pith2026250520962,
  author       = {Pith},
  title        = {Pith review of: Object-Centric Action-Enhanced Representations for Robot Visuo-Motor Policy Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OEYEOAO6}},
  note         = {Machine review of arXiv:2505.20962}
}
read the original abstract

Learning visual representations from observing actions to benefit robot visuo-motor policy generation is a promising direction that closely resembles human cognitive function and perception. Motivated by this, and further inspired by psychological theories suggesting that humans process scenes in an object-based fashion, we propose an object-centric encoder that performs semantic segmentation and visual representation generation in a coupled manner, unlike other works, which treat these as separate processes. To achieve this, we leverage the Slot Attention mechanism and use the SOLV model, pretrained in large out-of-domain datasets, to bootstrap fine-tuning on human action video data. Through simulated robotic tasks, we demonstrate that visual representations can enhance reinforcement and imitation learning training, highlighting the effectiveness of our integrated approach for semantic segmentation and encoding. Furthermore, we show that exploiting models pretrained on out-of-domain datasets can benefit this process, and that fine-tuning on datasets depicting human actions -- although still out-of-domain -- , can significantly improve performance due to close alignment with robotic tasks. These findings show the capability to reduce reliance on annotated or robot-specific action datasets and the potential to build on existing visual encoders to accelerate training and improve generalizability.

Figures

Figures reproduced from arXiv: 2505.20962 by the authors.

Figure 1
Figure 1. Approach Overview: Our proposed encoder extracts “what” and “where” visual object-centric embeddings, which are combined to form scene representations. These embeddings distill action-based knowledge acquired from pretraining on human action videos to effectively guide robot policy learning. are able to learn how things can be manipulated, e.g. moved, squeezed, stacked, and others. Although all senses participate in… view at source ↗
Figure 2
Figure 2. Our proposed architecture: Image embeddings produced by DINOv2 are processed by the Slot Attention Module to generate slots. The Slot Merger then outputs four final object-specific features, referred to as “what” vectors, each corresponding to a distinct semantic region. These are combined with the associated “where” vectors, which encode spatial information derived from the attention masks, to form a unified object… view at source ↗
Figure 3
Figure 3. Experimental Environment: We simulate the Franka Emika Panda robot arm pouring task within the TOTO framework. The goal is to transfer as many spheres as possible from the cup to the container. iteratively by attending to the input features, allowing it to bind to a specific object or semantic region, thus producing object-centric representations without supervision. SOLV [1] leverages the slot attention mechanism i… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Simulation frames from the TOTO pouring task, segmented by the Slot Attention masks of the SOLV model, with the Slot Merger module outputting 4, 6, and 8 slots respectively (from left to right) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Success rates and mean rewards of different number of slots configurations in the TOTO pouring simulation task. Each variant was trained five [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Success rates and mean rewards comparing “what” and unified “what” + “where” representations. Each configuration was trained five times and [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [1]

    Self-supervised Object-Centric Learning for Videos,

    G. Aydemir, W. Xie, and F. Guney, “Self-supervised Object-Centric Learning for Videos,” in Proc. of Advances in Neural Information Processing Systems, vol. 36, New Orleans (LA), USA, Dec. 2023, pp. 32 879–32 899

    work page 2023

  2. [2]

    J. J. Gibson, The Senses Considered as Perceptual Systems . Boston (MA), USA: Houghton Mifflin, 1966

    work page 1966

  3. [3]

    Building an Affordances Map with Interactive Perception,

    L. K. Le Goff, O. Yaakoubi, A. Coninx, and S. Doncieux, “Building an Affordances Map with Interactive Perception,” Frontiers in Neuro- robotics, vol. 16, p. 504459, May 2022

    work page 2022

  4. [4]

    VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training,

    Y . J. Ma, S. Sodhani, D. Jayaraman, O. Bastani, V . Kumar, and A. Zhang, “VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training,” arXiv:2210.00030, Sep. 2022

    arXiv 2022

  5. [5]

    DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control,

    Z. Cui, H. Pan, A. Iyer, S. Haldar, and L. Pinto, “DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control,” in Proc. of Advances in Neural Information Processing Systems , Vancouver (BC), Canada, Dec. 2024, pp. 33 933–33 961

    work page 2024

  6. [6]

    Ego4D: Around the World in 3,000 Hours of Egocentric Video,

    K. Grauman, A. Westbury, E. Byrne, Z. Chavis, A. Furnari, R. Girdhar, J. Hamburger, H. Jiang, M. Liu, X. Liu et al. , “Ego4D: Around the World in 3,000 Hours of Egocentric Video,” in Proc. of IEEE/CVF Conf. on Computer Vision and Pattern Recognition , New Orleans (LA), USA, Jun. 2022, pp. 18 995–19 012

    work page 2022

  7. [7]

    The ”something something

    R. Goyal, S. Ebrahimi Kahou, V . Michalski, J. Materzynska, S. West- phal, H. Kim, V . Haenel, I. Fruend, P. Yianilos, M. Mueller-Freitag, F. Hoppe, C. Thurau, I. Bax, and R. Memisevic, “The ”something something” video database for learning and evaluating visual common sense,” in Proc. of IEEE Intl. Conf. on Computer Vision , Venice, Italy, Oct. 2017, pp....

    work page 2017

  8. [8]

    ImageNet: A Large-Scale Hierarchical Image Database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” in Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, Miami (FL), USA, Jun. 2009, pp. 248–255

    work page 2009

Show all 28 references
  1. [9]

    RoboNet: Large-Scale Multi-Robot Learning,

    S. Dasari, F. Ebert, S. Tian, S. Nair, B. Bucher, K. Schmeckpeper, S. Singh, S. Levine, and C. Finn, “RoboNet: Large-Scale Multi-Robot Learning,” arXiv:1910.11215, Oct. 2019

    1910 arXiv

  2. [10]

    Iosifidis and A

    A. Iosifidis and A. Tefas, Deep Learning for Robot Perception and Cognition. Cambridge (MA), USA: Academic Press, 2022

    2022

  3. [11]

    Learning to Navigate Using Mid-Level Visual Priors,

    A. Sax, J. O. Zhang, B. Emi, A. Zamir, S. Savarese, L. Guibas, and J. Malik, “Learning to Navigate Using Mid-Level Visual Priors,” arXiv:1912.11121, Dec. 2019

    1912 arXiv

  4. [12]

    Train Offline, Test Online: A Real Robot Learning Benchmark,

    G. Zhou, V . Dean, M. K. Srirama, A. Rajeswaran, J. Pari, K. Hatch, A. Jain, T. Yu, P. Abbeel, L. Pinto, C. Finn, and A. Gupta, “Train Offline, Test Online: A Real Robot Learning Benchmark,” in Proc. of IEEE Intl. Conf. on Robotics and Automation , London, UK, May-Jun. 2023, p...

    2023

  5. [13]

    S. E. Palmer, Vision Science: Photons to Phenomenology. Cambridge (MA), USA: MIT press, 1999

    1999

  6. [14]

    Word-object and action-object association learning across early development,

    S. F. Eiteljoerge, M. Adam, B. Elsner, and N. Mani, “Word-object and action-object association learning across early development,” PLOS One, vol. 14, no. 8, p. e0220317, Aug. 2019

    2019

  7. [15]

    VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors,

    Y . Zhu, A. Joshi, P. Stone, and Y . Zhu, “VIOLA: Imitation Learning for Vision-Based Manipulation with Object Proposal Priors,” in Proc. of Conf. on Robot Learning, Atlanta (GA), USA, Nov. 2023, pp. 1199– 1210

    2023

  8. [16]

    Composing Pre- Trained Object-Centric Representations for Robotics From “What

    J. Shi, J. Qian, Y . J. Ma, and D. Jayaraman, “Composing Pre- Trained Object-Centric Representations for Robotics From “What” and “Where” Foundation Models,” in Proc. of IEEE Intl. Conf. on Robotics and Automation, Yokohama, Japan, May 2024, pp. 15 424–15 432

    2024

  9. [17]

    Object- Centric Learning with Slot Attention,

    F. Locatello, D. Weissenborn, T. Unterthiner, A. Mahendran, G. Heigold, J. Uszkoreit, A. Dosovitskiy, and T. Kipf, “Object- Centric Learning with Slot Attention,” in Proc. of Advances in Neural Information Processing Systems , vol. 33, Virtual, Dec. 2020, pp. 11 525–11 538

    2020

  10. [18]

    Invariant slot attention: Object discovery with slot- centric reference frames,

    O. Biza, S. Van Steenkiste, M. S. Sajjadi, G. F. Elsayed, A. Mahendran, and T. Kipf, “Invariant slot attention: Object discovery with slot- centric reference frames,” in Proc. of Intl. Conf. on Machine Learning, Honolulu (HI), USA, Jul. 2023, pp. 2507–2527

    2023

  11. [19]

    DINOv2: Learning Robust Visual Features without Supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khali- dov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A...

    2023 arXiv

  12. [20]

    MuJoCo: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “MuJoCo: A physics engine for model-based control,” in Proc. of IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems , Vilamoura, Portugal, Oct. 2012

    2012

  13. [21]

    The Franka Emika Robot: A Reference Platform for Robotics Research and Education,

    S. Haddadin, S. Parusel, L. Johannsmeier, S. Golz, S. Gabl, F. Walch, M. Sabaghian, C. J ¨ahne, L. Hausperger, and S. Haddadin, “The Franka Emika Robot: A Reference Platform for Robotics Research and Education,” IEEE Robotics & Automation Magazine , vol. 29, no. 2, pp. 46–64, ...

    2022

  14. [22]

    Behavioral Cloning from Observation,

    F. Torabi, G. Warnell, and P. Stone, “Behavioral Cloning from Observation,” in Proc. of Intl. Joint Conf. on Artificial Intelligence , Stockholm, Sweden, Jul. 2018, pp. 4950–4957

    2018

  15. [23]

    Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning,

    J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. D. Guo, M. Ghesh- laghi Azar, B. Piot, M. Valko, and R. Munos, “Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning,” in Proc. of Advances in Neural...

    2020

  16. [24]

    Learning Transferable Visual Models from Natural Language Supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning Transferable Visual Models from Natural Language Supervision,” in Proc. of Intl. Conf. on Machine Learning , Virtual, Jul. 2021, pp. 8748–8763

    2021

  17. [25]

    Momentum Con- trast for Unsupervised Visual Representation Learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum Con- trast for Unsupervised Visual Representation Learning,” in Proc. of IEEE/CVF Conf. on Computer Vision and Pattern Recognition, Seattle (W A), USA, Jun. 2020, pp. 9729–9738

    2020

  18. [26]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, Las Vegas (NV), USA, Jun. 2016, pp. 770–778

    2016

  19. [27]

    Video Instance Segmentation,

    L. Yang, Y . Fan, and N. Xu, “Video Instance Segmentation,” in Proc. of IEEE/CVF Intl. Conf. on Computer Vision , Seoul, Korea, October 2019, pp. 5188–5197

    2019

  20. [28]

    Offline Reinforcement Learning with Implicit Q-Learning,

    I. Kostrikov, A. Nair, and S. Levine, “Offline Reinforcement Learning with Implicit Q-Learning,” arXiv:2110.06169, Oct. 2021

    2021 arXiv

Pith tools

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