pith. machine review for the scientific record. sign in

arxiv: 2604.16440 · v1 · submitted 2026-04-07 · 💻 cs.RO · cs.AI

Recognition: 2 theorem links

· Lean Theorem

LatentMimic: Terrain-Adaptive Locomotion via Latent Space Imitation

Zhiquan Wang , Yunyu Liu , Dipam Patel , Ayush Kumar , Aniket Bera , Bedrich Benes
Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:39 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords quadruped locomotionimitation learningterrain adaptationlatent spacemotion capturerobot controlstyle preservation
0
0 comments X

The pith

Minimizing marginal latent divergence from motion capture priors enables quadruped robots to adapt to irregular terrains while preserving original locomotion styles.

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

Existing imitation methods for quadruped robots create a direct conflict: following motion capture data too closely prevents the geometric changes needed for rough terrain, while terrain-focused policies erode the desired walking style. LatentMimic resolves this by shifting the objective to latent space, where the policy only needs to match the overall distribution of states and actions from the motion prior rather than every pose exactly. This conditional relaxation keeps the core gait pattern intact but frees the end-effectors to adjust foot placements for surface irregularities. A separate terrain adaptation module with a dynamic replay buffer further counters the policy shifts that occur when moving between different ground types. Experiments across four distinct locomotion styles and four terrains show higher traversal success than prior motion-tracking approaches without loss of stylistic quality.

Core claim

By minimizing the marginal latent divergence between the policy's state-action distribution and a learned mocap prior, LatentMimic supplies a conditional relaxation of rigid pose-tracking objectives. This formulation preserves gait topology while permitting independent end-effector adaptations for irregular terrains. A terrain adaptation module equipped with a dynamic replay buffer resolves the policy's distribution shifts across different terrains, producing higher terrain traversal success rates than state-of-the-art motion-tracking methods while retaining high stylistic fidelity across four locomotion styles and four terrains.

What carries the argument

Marginal latent divergence minimization between the policy state-action distribution and the learned mocap prior, which relaxes strict pose tracking to allow terrain-specific end-effector adjustments while keeping gait topology fixed.

If this is right

  • Quadruped robots achieve higher success rates when crossing irregular terrain compared with rigid motion-tracking controllers.
  • Stylistic fidelity remains high across multiple distinct locomotion styles without requiring per-terrain retraining.
  • Policy distribution shifts induced by new terrain surfaces are mitigated by the dynamic replay buffer.
  • End-effector positions can vary independently while the core gait sequence stays consistent with the motion prior.

Where Pith is reading between the lines

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

  • The same distribution-matching approach in latent space could extend to other control domains where style preservation conflicts with task-specific adaptation, such as arm manipulation.
  • Testing on continuously changing terrain during a single episode would reveal whether the replay buffer mechanism scales beyond discrete terrain switches.
  • If the latent prior is learned from a broader set of motion data, the method might support seamless transitions between styles without explicit conditioning.

Load-bearing premise

Minimizing marginal latent divergence will preserve overall gait topology and stylistic control even when end-effectors make independent adjustments for terrain irregularities.

What would settle it

Run the trained policy on a previously unseen terrain type, such as loose gravel or steep inclines, and measure whether traversal success rate falls below that of direct motion-tracking baselines or whether quantitative style metrics decline.

Figures

Figures reproduced from arXiv: 2604.16440 by Aniket Bera, Ayush Kumar, Bedrich Benes, Dipam Patel, Yunyu Liu, Zhiquan Wang.

Figure 1
Figure 1. Figure 1: Left to right: Stair terrain, Noisy terrain, Wavy terrain. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of LatentMimic: a) We first pretrain a motion encoder to encode the motion transitions into a latent space [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The input of our policy consists of proprioceptive [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: A target motion predicted from flat terrain data may [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: The latent mimic reward: both the simulated and the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Terrain Adaptation Module: the current policy [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The figure shows the t-SNE visualization of latent [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The left figure depicts a frame of pace in which [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
read the original abstract

Developing natural and diverse locomotion controllers for quadruped robots that can adapt to complex terrains while preserving motion style remains a significant challenge. Existing imitation-based methods face a fundamental optimization trade-off: strict adherence to motion capture (mocap) references penalizes the geometric deviations required for terrain adaptability, whereas terrain-centric policies often compromise stylistic fidelity. We introduce LatentMimic, a novel locomotion learning framework that decouples stylistic fidelity from geometric constraints. By minimizing the marginal latent divergence between the policy's state-action distribution and a learned mocap prior, our approach provides a conditional relaxation of rigid pose-tracking objectives. This formulation preserves gait topology while permitting independent end-effector adaptations for irregular terrains. We further introduce a terrain adaptation module with a dynamic replay buffer to resolve the policy's distribution shifts across different terrains. We validate our method across four locomotion styles and four terrains, demonstrating that LatentMimic enables effective terrain-adaptive locomotion, achieving higher terrain traversal success rates than state-of-the-art motion-tracking methods while maintaining high stylistic fidelity.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript introduces LatentMimic, a locomotion learning framework for quadruped robots that decouples stylistic fidelity from geometric constraints by minimizing the marginal latent divergence between the policy's state-action distribution and a learned mocap prior. This provides a conditional relaxation of rigid pose-tracking objectives while preserving gait topology and permitting end-effector adaptations for irregular terrains. A terrain adaptation module with a dynamic replay buffer is introduced to address distribution shifts. The approach is validated across four locomotion styles and four terrains, claiming higher terrain traversal success rates than state-of-the-art motion-tracking methods while maintaining high stylistic fidelity.

Significance. If the empirical claims hold under rigorous validation, this work could meaningfully advance imitation-based control for legged robots by offering a latent-space mechanism to relax strict mocap tracking without sacrificing stylistic control, addressing a persistent optimization trade-off in the field. The marginal divergence formulation and dynamic replay buffer represent potentially reusable ideas for terrain-adaptive policies.

major comments (2)
  1. [Abstract] Abstract: the claim of 'higher terrain traversal success rates than state-of-the-art motion-tracking methods' supplies no quantitative metrics, error bars, ablation details, or experimental protocol, so the data-to-claim link cannot be evaluated.
  2. [Method description] Method (marginal latent divergence objective): the central assumption that minimizing marginal (not conditional) latent divergence between policy state-action distribution and the mocap prior preserves gait topology while allowing independent end-effector adaptations is not shown to hold; if the prior is learned on flat-ground mocap, terrain-induced foot-placement changes can produce averaged or collapsed gaits that lose periodic footfall structure, since the objective averages over the joint distribution without penalizing conditional mismatches (e.g., terrain height map paired with phase or velocity).
minor comments (1)
  1. [Abstract] Abstract: the four locomotion styles and four terrains are referenced but not named or characterized, which would improve clarity for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We address each major comment below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'higher terrain traversal success rates than state-of-the-art motion-tracking methods' supplies no quantitative metrics, error bars, ablation details, or experimental protocol, so the data-to-claim link cannot be evaluated.

    Authors: We agree that the abstract would benefit from greater specificity to support the claims. In the revised version we will update the abstract to report concrete quantitative results, including average terrain traversal success rates with standard deviations across repeated trials, the number of evaluation episodes per condition, and a brief reference to the experimental protocol used for the four styles and four terrains. revision: yes

  2. Referee: [Method description] Method (marginal latent divergence objective): the central assumption that minimizing marginal (not conditional) latent divergence between policy state-action distribution and the mocap prior preserves gait topology while allowing independent end-effector adaptations is not shown to hold; if the prior is learned on flat-ground mocap, terrain-induced foot-placement changes can produce averaged or collapsed gaits that lose periodic footfall structure, since the objective averages over the joint distribution without penalizing conditional mismatches (e.g., terrain height map paired with phase or velocity).

    Authors: We acknowledge the concern that the marginal formulation could in principle average over conditional mismatches. Our current empirical results (Section 5) show that gait topology is preserved across terrains, as evidenced by consistent footfall periodicity and style metrics. Nevertheless, we agree that an explicit demonstration of why the marginal objective does not induce collapse is needed. In the revision we will add a new subsection with (i) footfall timing consistency metrics under varying terrain heights, (ii) latent-space trajectory visualizations conditioned on terrain inputs, and (iii) a discussion of how the terrain-conditioned policy and dynamic replay buffer prevent averaging by enabling local adaptations while the marginal term only regularizes global style. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained and empirically grounded.

full rationale

The paper introduces LatentMimic via an explicit modeling choice: minimizing marginal latent divergence between the policy state-action distribution and a separately learned mocap prior, then augments it with a terrain adaptation module and dynamic replay buffer. This is presented as a design decision whose benefits (preserved gait topology with terrain adaptability) are tested through experiments on four styles and four terrains, reporting higher success rates than baselines. No load-bearing claim reduces to a self-citation chain, a fitted parameter renamed as a prediction, or definitional equivalence; the mocap prior is learned from motion data independent of the terrain-specific evaluation metrics. The central formulation is therefore not forced by its own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

Review is limited to the abstract; no explicit free parameters, axioms, or invented entities are enumerated in the provided text. The framework implicitly relies on the existence of a learnable mocap prior and the validity of latent-space divergence as a proxy for style.

invented entities (2)
  • marginal latent divergence objective no independent evidence
    purpose: To relax rigid pose tracking while preserving gait topology
    Introduced as the core mechanism for decoupling style from geometry.
  • dynamic replay buffer no independent evidence
    purpose: To mitigate policy distribution shifts across terrains
    Added to stabilize learning when terrain changes.

pith-pipeline@v0.9.0 · 5489 in / 1379 out tokens · 57203 ms · 2026-05-10T19:39:39.983949+00:00 · methodology

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

44 extracted references · 3 canonical work pages

  1. [1]

    Gait and trajectory optimization for legged systems through phase-based end- effector parameterization,

    A. W. Winkler, C. D. Bellicoso, M. Hutter, and J. Buchli, “Gait and trajectory optimization for legged systems through phase-based end- effector parameterization,”IEEE Robot. Autom. Lett., vol. 3, pp. 1560– 1567, July 2018

    2018

  2. [2]

    Dynamically diverse legged locomotion for rough terrain,

    K. Byl and R. Tedrake, “Dynamically diverse legged locomotion for rough terrain,” in2009 IEEE International Conference on Robotics and Automation, pp. 1607–1608, IEEE, May 2009

    2009

  3. [3]

    A linearization of centroidal dynamics for the model-predictive control of quadruped robots,

    W. Chi, X. Jiang, and Y . Zheng, “A linearization of centroidal dynamics for the model-predictive control of quadruped robots,” in 2022 International Conference on Robotics and Automation (ICRA), pp. 4656–4663, IEEE, May 2022

    2022

  4. [4]

    Design of HyQ – a hydraulically and electrically actuated quadruped robot,

    C. Semini, N. G. Tsagarakis, E. Guglielmino, M. Focchi, F. Cannella, and D. G. Caldwell, “Design of HyQ – a hydraulically and electrically actuated quadruped robot,”Proc Inst Mech Eng Part I J Syst Control Eng, vol. 225, pp. 831–849, Sept. 2011

    2011

  5. [5]

    A factor-graph approach for optimization problems with dynamics constraints,

    M. Xie, A. Escontrela, and F. Dellaert, “A factor-graph approach for optimization problems with dynamics constraints,”arXiv, Nov. 2020

    2020

  6. [6]

    Dynamic locomotion and whole-body control for quadrupedal robots,

    C. Dario Bellicoso, F. Jenelten, P. Fankhauser, C. Gehring, J. Hwangbo, and M. Hutter, “Dynamic locomotion and whole-body control for quadrupedal robots,” in2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3359–3365, IEEE, Sept. 2017

    2017

  7. [7]

    Learning to walk in minutes using massively parallel deep reinforcement learning,

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter, “Learning to walk in minutes using massively parallel deep reinforcement learning,”arXiv [cs.RO], Sept. 2021

    2021

  8. [8]

    Rapid locomotion via reinforcement learning,

    G. B. Margolis, G. Yang, K. Paigwar, T. Chen, and P. Agrawal, “Rapid locomotion via reinforcement learning,”arXiv [cs.RO], May 2022

    2022

  9. [9]

    Learning agile and dynamic motor skills for legged robots,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots,”Sci. Robot., vol. 4, p. eaau5872, Jan. 2019

    2019

  10. [10]

    Emergence of locomotion behaviours in rich environments,

    N. Heess, D. Tb, S. Sriram, J. Lemmon, J. Merel, G. Wayne, Y . Tassa, T. Erez, Z. Wang, S. M. A. Eslami, M. Riedmiller, and D. Silver, “Emergence of locomotion behaviours in rich environments,”arXiv [cs.AI], July 2017

    2017

  11. [11]

    Latent conditioned loco-manipulation using motion priors,

    M. St˛ epie´n, R. Kourdis, C. Roux, and O. Stasse, “Latent conditioned loco-manipulation using motion priors,” in2025 IEEE-RAS 24th International Conference on Humanoid Robots, IEEE, 2025

    2025

  12. [12]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,”Science robotics, vol. 5, no. 47, p. eabc5986, 2020

    2020

  13. [13]

    Trans: Terrain- aware reinforcement learning for agile navigation of quadruped robots under social interactions,

    W. Zhu, I. T. Kurniawan, Y . Zhao, and M. Hayashibe, “Trans: Terrain- aware reinforcement learning for agile navigation of quadruped robots under social interactions,” 2026

    2026

  14. [14]

    Training and simulation of quadrupedal robot in adaptive stair climbing for indoor firefighting: An end-to-end reinforcement learning approach,

    B. Huang, B. Huang, and Y . Hou, “Training and simulation of quadrupedal robot in adaptive stair climbing for indoor firefighting: An end-to-end reinforcement learning approach,” 2026

    2026

  15. [15]

    Learning agile robotic locomotion skills by imitating animals,

    X. B. Peng, E. Coumans, T. Zhang, T.-W. Lee, J. Tan, and S. Levine, “Learning agile robotic locomotion skills by imitating animals,”arXiv [cs.RO], Apr. 2020

    2020

  16. [16]

    Learning terrain-adaptive locomotion with agile behaviors by imitating animals,

    T. Li, Y . Zhang, C. Zhang, Q. Zhu, J. Sheng, W. Chi, C. Zhou, and L. Han, “Learning terrain-adaptive locomotion with agile behaviors by imitating animals,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, Oct. 2023

    2023

  17. [17]

    Mode-adaptive neural networks for quadruped motion control,

    H. Zhang, S. Starke, T. Komura, and J. Saito, “Mode-adaptive neural networks for quadruped motion control,”ACM Trans. Graph., vol. 37, pp. 1–11, Aug. 2018

    2018

  18. [18]

    Learning robust perceptive locomotion for quadrupedal robots in the wild,

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning robust perceptive locomotion for quadrupedal robots in the wild,”Science robotics, vol. 7, no. 62, p. eabk2822, 2022

    2022

  19. [19]

    Rma: Rapid motor adaptation for legged robots,

    A. Kumar, Z. Fu, D. Pathak, and J. Malik, “Rma: Rapid motor adaptation for legged robots,” inRobotics: Science and Systems XVII, RSS2021, Robotics: Science and Systems Foundation, July 2021

    2021

  20. [20]

    Sim-to-real learning of all common bipedal gaits via periodic reward composition,

    J. Siekmann, Y . Godse, A. Fern, and J. Hurst, “Sim-to-real learning of all common bipedal gaits via periodic reward composition,” in2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 7309–7315, IEEE, 2021

    2021

  21. [21]

    Robust recovery controller for a quadrupedal robot using deep reinforcement learning,

    J. Lee, J. Hwangbo, and M. Hutter, “Robust recovery controller for a quadrupedal robot using deep reinforcement learning,” 2019

    2019

  22. [22]

    Teacher–student curriculum learning,

    T. Matiisen, A. Oliver, T. Cohen, and J. Schulman, “Teacher–student curriculum learning,”IEEE Transactions on Neural Networks and Learning Systems, vol. 31, p. 3732–3740, Sept. 2020

    2020

  23. [23]

    Real-time imitation of human whole-body motions by humanoids,

    J. Koenemann, F. Burget, and M. Bennewitz, “Real-time imitation of human whole-body motions by humanoids,” in2014 IEEE Inter- national Conference on Robotics and Automation (ICRA), pp. 2806– 2812, IEEE, 2014

    2014

  24. [24]

    Deepmimic: example-guided deep reinforcement learning of physics-based charac- ter skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. van de Panne, “Deepmimic: example-guided deep reinforcement learning of physics-based charac- ter skills,”ACM Transactions on Graphics, July 2018

    2018

  25. [25]

    Learning multi-skill legged loco- motion using conditional adversarial motion priors,

    N. Huang, Z. Xie, and Q. Li, “Learning multi-skill legged loco- motion using conditional adversarial motion priors,”arXiv preprint arXiv:2509.21810, 2025

    work page arXiv 2025

  26. [26]

    AMP: Adversarial motion priors for stylized physics-based character con- trol,

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “AMP: Adversarial motion priors for stylized physics-based character con- trol,”arXiv [cs.GR], Apr. 2021

    2021

  27. [27]

    Generative adversarial imitation learning,

    J. Ho and S. Ermon, “Generative adversarial imitation learning,” 2016

    2016

  28. [28]

    Adversarial motion priors make good substitutes for complex reward functions,

    A. Escontrela, X. B. Peng, W. Yu, T. Zhang, A. Iscen, K. Goldberg, and P. Abbeel, “Adversarial motion priors make good substitutes for complex reward functions,” 2022

    2022

  29. [29]

    Adversarial motion priors make good substitutes for complex reward functions,

    A. Escontrela, X. B. Peng, W. Yu, T. Zhang, A. Iscen, K. Goldberg, and P. Abbeel, “Adversarial motion priors make good substitutes for complex reward functions,” in2022 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), IEEE, Oct. 2022

    2022

  30. [30]

    Learning robust and agile legged locomotion using adversarial motion priors,

    J. Wu, G. Xin, C. Qi, and Y . Xue, “Learning robust and agile legged locomotion using adversarial motion priors,”IEEE Robot. Autom. Lett., vol. 8, pp. 4975–4982, Aug. 2023

    2023

  31. [31]

    Generalized animal imitator: Agile locomotion with versatile motion prior,

    R. Yang, Z. Chen, J. Ma, C. Zheng, Y . Chen, Q. Nguyen, and X. Wang, “Generalized animal imitator: Agile locomotion with versatile motion prior,”arXiv [cs.RO], Oct. 2023

    2023

  32. [32]

    Learning multiple gaits within latent space for quadruped robots,

    J. Wu, Y . Xue, and C. Qi, “Learning multiple gaits within latent space for quadruped robots,”arXiv [cs.RO], Aug. 2023

    2023

  33. [33]

    Bcamp: A behavior- controllable motion control method based on adversarial motion priors for quadruped robot,

    Y . Peng, Z. Cai, L. Zhang, and X. Wang, “Bcamp: A behavior- controllable motion control method based on adversarial motion priors for quadruped robot,”Applied Sciences, vol. 15, no. 6, p. 3356, 2025

    2025

  34. [34]

    Terrain- aware quadrupedal locomotion via reinforcement learning,

    H. Shi, Q. Zhu, L. Han, W. Chi, T. Li, and M. Q.-H. Meng, “Terrain- aware quadrupedal locomotion via reinforcement learning,”arXiv preprint arXiv:2310.04675, 2023

    work page arXiv 2023

  35. [35]

    Infogan: Interpretable representation learning by infor- mation maximizing generative adversarial nets,

    X. Chen, Y . Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel, “Infogan: Interpretable representation learning by infor- mation maximizing generative adversarial nets,” 2016

    2016

  36. [36]

    Deep generative models with learnable knowledge constraints,

    Z. Hu, Z. Yang, R. Salakhutdinov, and X. L. etc, “Deep generative models with learnable knowledge constraints,” 2018

    2018

  37. [37]

    Stability analysis and generalization bounds of adversarial training,

    J. Xiao, Y . Fan, R. Sun, J. Wang, and Z.-Q. Luo, “Stability analysis and generalization bounds of adversarial training,” 2022

    2022

  38. [38]

    f-gan: Training generative neural samplers using variational divergence minimization,

    S. Nowozin, B. Cseke, and R. Tomioka, “f-gan: Training generative neural samplers using variational divergence minimization,” 2016

    2016

  39. [39]

    Retargetting motion to new characters,

    M. Gleicher, “Retargetting motion to new characters,” inProceedings of the 25th annual conference on Computer graphics and interactive techniques - SIGGRAPH ’98, ACM Press, 1998

    1998

  40. [40]

    Terrain-adaptive locomotion skills using deep reinforcement learning,

    X. B. Peng, G. Berseth, and M. van de Panne, “Terrain-adaptive locomotion skills using deep reinforcement learning,”ACM Trans. Graph., vol. 35, pp. 1–12, July 2016

    2016

  41. [41]

    Learning and transfer of modulated locomotor controllers,

    N. Heess, G. Wayne, Y . Tassa, T. Lillicrap, M. Riedmiller, and D. Silver, “Learning and transfer of modulated locomotor controllers,” arXiv preprint arXiv:1610.05182, 2016

    work page arXiv 2016

  42. [42]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017

    2017

  43. [43]

    Isaac gym: High performance gpu-based physics simulation for robot learning,

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Mack- lin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac gym: High performance gpu-based physics simulation for robot learning,” 2021

    2021

  44. [44]

    Fast and accurate deep network learning by exponential linear units (elus),

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (elus),” 2016. 8

    2016