Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Standard sparse autoencoders break temporal coherence in video features; contrastive strength is a controllable knob that restores it and can beat raw features on action and retrieval.

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 Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 12:00 UTC

load-bearing objection Solid first systematic video SAE study: TopK temporal flicker is real, contrastive strength is a usable Pareto knob, and the CLIP monosemanticity control is a genuine contribution. the 3 major comments →

arxiv 2604.03919 v1 submitted 2026-04-05 cs.CV cs.AI

Interpreting Video Representations with Spatio-Temporal Sparse Autoencoders

Atahan Dokme , Sriram Vishwanath This is my paper
classification cs.CV cs.AI
keywords Sparse AutoencodersVideo UnderstandingMechanistic InterpretabilityTemporal CoherenceFeature DecompositionContrastive LearningAction ClassificationText-Video Retrieval
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.

This paper is the first systematic study of sparse autoencoders on video representations from frozen vision backbones. It shows that ordinary TopK SAEs do decompose video patches into interpretable, monosemantic concepts, but hard top-k selection makes feature assignments flicker across frames and cuts lag-1 autocorrelation by about 36 percent. The authors add spatio-temporal contrastive losses and Matryoshka hierarchical grouping so that the same semantic features stay active over time. The weight and temperature of the contrastive term set a smooth trade-off between reconstruction fidelity and temporal coherence, and different settings win on reconstruction, coherence, action discrimination, or interpretability. On two backbones and two datasets, the best contrastive configurations improve linear action probes by up to +3.9 points over raw features and text-video retrieval by up to 2.8 times R@1, while causal ablations show that the predictive signal concentrates into a small identifiable set of features.

Core claim

Hard TopK sparse autoencoders produce monosemantic video features but destroy temporal coherence; spatio-temporal contrastive objectives and Matryoshka grouping recover and can exceed raw lag-1 autocorrelation, and the contrastive strength (via lambda or temperature) is a single controllable axis that places each variant on a reconstruction-coherence Pareto frontier while improving action discrimination and text-video retrieval.

What carries the argument

Spatio-temporal contrastive SAE variants (Temporal, Separate, Raster) plus optional Matryoshka hierarchical grouping: InfoNCE losses encourage consistent sparse activations between temporally or spatially adjacent patches, competing with reconstruction for limited TopK capacity so that the contrastive weight sets the operating point.

Load-bearing premise

The main temporal loss treats the same fixed spatial patch index across consecutive frames as a true positive pair, which assumes the camera and scene keep that location roughly corresponding from frame to frame.

What would settle it

On a large set of clips with strong camera motion or object tracking failure, re-run Temporal and Temporal+M SAEs against a version that forms positives from optical-flow or object-tracked correspondences: if lag-1, probe accuracy, and retrieval gains vanish or reverse under fixed-index pairing but return under true correspondence, the central trade-off claim fails for general video.

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

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

3 major / 5 minor

Summary. The paper presents the first systematic study of TopK Sparse Autoencoders applied to video representations from DINOv2 and VideoMAE on SSv2 and Kinetics-400. It shows that standard SAEs yield monosemantic features but destroy temporal coherence (lag-1 autocorrelation falls ~36% via hard TopK). The authors introduce three spatio-temporal contrastive variants (Temporal, Separate, Raster-scan) plus Matryoshka hierarchical grouping, and demonstrate via a 29-configuration λ/τ ablation that contrastive strength places models on a reconstruction–coherence Pareto frontier. Temporal+M can exceed raw DINOv2 lag-1 (0.462 vs 0.435) while improving frozen linear action probes by up to +3.9% and text-video retrieval by up to 2.8× R@1 (with a 768-d reconstructed control). Cross-backbone MS evaluation under CLIP reveals a DINOv2-alignment artifact; causal ablation (no probe retrain) shows contrastive training concentrates predictive signal into few identifiable features. Soft-sparsity, frozen-decoder, and post-hoc smoothing controls further isolate the mechanism.

Significance. If the results hold, this is a useful first map of how SAEs behave on video features and a practical design principle: contrastive strength is a controllable axis rather than a fixed hyperparameter. The multi-seed 29-config ablation, two-backbone/two-dataset design, post-hoc baselines that improve lag-1 but collapse probe accuracy, frozen-decoder control, dimensionality-matched retrieval control, and causal ablation without retraining are concrete strengths that make the Pareto and concentration claims falsifiable and reproducible. The monosemanticity artifact finding is a methodological contribution of independent interest for vision SAE evaluation. Absolute gains remain modest (frozen probes, 10K-clip scale), so the work is foundational rather than transformative, but it supplies clear operating points (Raster for reconstruction/MS, Temporal+M for coherence, etc.) that subsequent video interpretability work can build on.

major comments (3)
  1. §3.2 Eq. (4) and §7: temporal positives are formed by fixed spatial-index pairing across consecutive frames. The paper correctly scopes this to stationary-camera SSv2 and notes weaker K400 gains, but the central claim that contrastive strength recovers/exceeds temporal coherence is still partly conditioned on approximate spatial correspondence. A quantitative breakdown of lag-1 / probe gains stratified by estimated camera motion (or a small optical-flow pairing ablation on a K400 subset) would make the scope of the Pareto frontier explicit rather than qualitative.
  2. Table 3 / §6.1 and Table 5: absolute probe accuracies remain low (raw DINOv2 17.4% → best SAE 20.4% on SSv2; supervised DINOv2 is ~45%). The relative +3.9% and 2.8× R@1 claims are supported by the 768-d reconstructed control and multi-seed variance <1%, but the manuscript should more prominently state that these are frozen linear-probe / Ridge-projection gains at 10K-clip scale, not end-to-end video understanding improvements, so readers do not over-read the absolute numbers.
  3. §5.1 Table 1 and MS protocol: the CLIP-neutral control cleanly shows the DINOv2-alignment artifact (MS ≈ 0.68 for both backbones). Because the paper still reports DINOv2-based MS for within-backbone comparisons in Table 3, it should either (a) report CLIP-MS alongside every MS number or (b) explicitly restrict all MS claims to within-backbone ranking only, to avoid residual cross-backbone misreading.
minor comments (5)
  1. Figure 2 caption and panels (c,d): the temporal flicker comparison is central; adding a quantitative per-frame activation trace or lag-1 callout on the figure itself would make the qualitative claim easier to verify.
  2. §4: training times and batch sizes are given; a short note on whether the 29-config sweep used the same three seeds for every (λ, τ) point (or only for the reported bests) would clarify variance claims.
  3. §6.5 layer analysis: the claim that temporal contrastive is most beneficial at the final DINOv2 layer (and harmful at layer 7) is important for practitioners; a one-sentence mechanistic speculation or pointer to the supplementary tables would help.
  4. Notation: L_temp in Eq. (4) uses N for both batch size and the sum index; a brief clarification that negatives are other patches in the batch (same and other clips) is already present but could be moved closer to the equation.
  5. References: T-SAE (Bhalla et al.) is correctly credited as the 1D precursor; a short sentence distinguishing the video setting (spatial + temporal adjacency, motion-induced content change) from token-level text would further situate the contribution.

Circularity Check

0 steps flagged

No significant circularity: contrastive training objectives are distinct from the external evaluation metrics (lag-1, probe, R@1, CLIP-MS), and claims rest on ablations and controls rather than definitional identity or self-citation chains.

full rationale

This is an empirical methods paper. The central claims (controllable reconstruction–coherence Pareto via λ/τ, Temporal+M lag-1 exceeding raw DINOv2, probe/retrieval gains, causal concentration) are established by training SAEs under stated losses (L_base + InfoNCE variants + optional Matryoshka) and measuring separate quantities: R², lag-1 autocorrelation of activations, frozen linear probes, Ridge text-video retrieval with a 768-d reconstructed dimensionality control, and probe-weight ablation without retraining. InfoNCE on fixed-index adjacent pairs encourages consistency but is not identical to lag-1 autocorrelation by construction; post-hoc EMA/Temporal-Union baselines improve lag-1 while collapsing probe accuracy, isolating the learned-objective benefit. The MS backbone-alignment finding is a control (CLIP-neutral MS collapses the DINOv2–VideoMAE gap), not a circular metric. Citations (T-SAE, BatchTopK/Matryoshka, TopK, Pach MS, Stevens ablation) are external prior work, not load-bearing self-uniqueness theorems. No fitted parameter is renamed as a prediction of a quantity forced by that fit; λ/τ sweeps report an observed trade-off. Fixed spatial-index pairing is a scoped assumption (SSv2 stationary cameras), not a definitional loop. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The work is empirical deep-learning methodology. It inherits standard SAE and InfoNCE machinery, freezes public backbones, and introduces contrastive pairing rules and a hierarchical split whose hyperparameters are chosen by ablation. No new physical entities; the free parameters are the usual training knobs that the central trade-off claim depends on.

free parameters (7)
  • contrastive weight λ (and spatial λ_s)
    Primary control of the reconstruction–coherence trade-off; swept over {0.01, 0.05, 0.5}; default λ_t=0.1, λ_s=0.05.
  • contrastive temperature τ
    Second control of contrastive strength; swept {0.05, 0.2, 0.5}; default 0.1; Temporal+M at τ=0.2 exceeds raw lag-1.
  • TopK sparsity k
    Fixed at 64 following Gao et al.; hard selection is identified as the source of temporal flicker.
  • dictionary expansion H / expansion factor
    Fixed at 8× (H=6144); capacity shared among reconstruction, contrastive, and Matryoshka objectives.
  • Matryoshka high-level fraction and α_mat
    Default 20/80 split with α_mat=0.1; other ratios ablated in supplement; concentrates action signal into 20% of dictionary.
  • auxiliary dead-latent coefficient α
    Fixed at 0.03; standard SAE hyperparameter affecting feature recovery.
  • training scale (clips / epochs)
    Main results on 10K clips, 10 epochs; larger scales deferred to supplement.
axioms (6)
  • domain assumption Hard TopK with straight-through estimator on values is a valid sparse encoder (Gao et al.).
    Used throughout Eqs. 1–3 and all variants; non-differentiable index selection is accepted as standard.
  • domain assumption InfoNCE on cosine similarity of sparse activation vectors encourages useful consistency without collapse.
    Eq. 4 and variants; negatives drawn from batch including same-clip different positions.
  • ad hoc to paper Fixed spatial-index pairing across consecutive frames is a valid positive pair for temporal contrast.
    Section 3.2; deliberate design choice that avoids external trackers but assumes limited camera motion.
  • domain assumption DINOv2 / VideoMAE final-layer (and selected intermediate) patch features are suitable SAE inputs.
    Section 4 experimental setup; frozen backbones, no end-to-end fine-tuning.
  • domain assumption Linear probes and Ridge projections onto CLIP text space measure action-discriminative and multimodal quality of SAE features.
    Sections 4, 6.3, 6.4; frozen-feature evaluation protocol.
  • domain assumption MonoSemanticity Score (activation-weighted pairwise similarity of top-activating clips) measures feature interpretability.
    Adopted from Pach et al.; paper then shows backbone-alignment artifact under CLIP control.
invented entities (3)
  • Raster-scan contrastive SAE (R-SAE) no independent evidence
    purpose: Unify spatial and temporal adjacency into one 1D contrastive loss via row-by-row then frame serialization.
    New pairing scheme relative to T-SAE; no independent evidence outside this paper’s ablations.
  • Separate spatio-temporal contrastive SAE (ST-SAE) no independent evidence
    purpose: Independent λ_t and λ_s control of temporal vs spatial coherence.
    Video-specific extension of temporal contrastive idea; evidence is internal ablation.
  • Matryoshka hierarchical grouping for video SAEs (+M) no independent evidence
    purpose: Concentrate action-relevant features into a compact high-level dictionary subset via auxiliary reconstruction.
    Applies BatchTopK Matryoshka (Bussmann et al.) to video SAEs; concentration claim supported by probe-on-high-level-only and causal ablation inside the paper.

pith-pipeline@v1.1.0-grok45 · 18901 in / 3850 out tokens · 27207 ms · 2026-07-13T12:00:15.349630+00:00 · methodology

0 comments
read the original abstract

We present the first systematic study of Sparse Autoencoders (SAEs) on video representations. Standard SAEs decompose video into interpretable, monosemantic features but destroy temporal coherence: hard TopK selection produces unstable feature assignments across frames, reducing autocorrelation by 36%. We propose spatio-temporal contrastive objectives and Matryoshka hierarchical grouping that recover and even exceed raw temporal coherence. The contrastive loss weight controls a tunable trade-off between reconstruction and temporal coherence. A systematic ablation on two backbones and two datasets shows that different configurations excel at different goals: reconstruction fidelity, temporal coherence, action discrimination, or interpretability. Contrastive SAE features improve action classification by +3.9% over raw features and text-video retrieval by up to 2.8xR@1. A cross-backbone analysis reveals that standard monosemanticity metrics contain a backbone-alignment artifact: both DINOv2 and VideoMAE produce equally monosemantic features under neutral (CLIP) similarity. Causal ablation confirms that contrastive training concentrates predictive signal into a small number of identifiable features.

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

29 extracted references · 1 canonical work pages

  1. [1]

    Usha Bhalla, Alex Oesterling, Claudio Mayrink Verdun, Himabindu Lakkaraju, and Flavio Calmon. 2026. Temporal Sparse Autoencoders: Leveraging the Sequen- tial Nature of Language for Interpretability. InThe Fourteenth International Confer- ence on Learning Representations. https://openreview.net/forum?id=bojVI4l9Kn

    2026

  2. [2]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. 2023. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread(2023)

    2023

  3. [3]

    Bart Bussmann, Patrick Leask, and Neel Nanda. 2024. BatchTopK Sparse Autoen- coders. arXiv:2412.06410 [cs.LG] https://arxiv.org/abs/2412.06410

    Pith/arXiv arXiv 2024

  4. [4]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey

  5. [5]

    arXiv:2309.08600 [cs.LG] https://arxiv.org/abs/2309.08600

    Sparse Autoencoders Find Highly Interpretable Features in Language Models. arXiv:2309.08600 [cs.LG] https://arxiv.org/abs/2309.08600

    Pith/arXiv arXiv

  6. [6]

    Ishan Dave, Rohit Gupta, Mamshad Nayeem Rizve, and Mubarak Shah. 2022. TCLR: Temporal contrastive learning for video representation.Computer Vision and Image Understanding219 (June 2022), 103406. doi:10.1016/j.cviu.2022.103406

    work page doi:10.1016/j.cviu.2022.103406 2022

  7. [7]

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. Scaling and evaluating sparse autoencoders. arXiv:2406.04093 [cs.LG] https://arxiv.org/abs/2406.04093

    Pith/arXiv arXiv 2024

  8. [8]

    something something

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzyńska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. 2017. The "something something" video database for learning and evaluating visual common sense. arXiv:1706.04261 [cs....

    Pith/arXiv arXiv 2017

  9. [9]

    Bear Häon, Kaylene Stocking, Ian Chuang, and Claire Tomlin. 2025. Mechanistic interpretability for steering vision-language-action models. arXiv:2509.00328 [cs.RO] https://arxiv.org/abs/2509.00328

    Pith/arXiv arXiv 2025

  10. [10]

    Sonia Joseph, Praneet Suresh, Ethan Goldfarb, Lorenz Hufe, Yossi Gandels- man, Robert Graham, Danilo Bzdok, Wojciech Samek, and Blake Aaron Richards. 2025. Steering CLIP’s vision transformer with sparse autoencoders. arXiv:2504.08729 [cs.CV] https://arxiv.org/abs/2504.08729

    Pith/arXiv arXiv 2025

  11. [11]

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. 2017. The Kinetics Human Action Video Dataset. arXiv:1705.06950 [cs.CV] https://arxiv.org/abs/1705.06950

    Pith/arXiv arXiv 2017

  12. [12]

    Anwar, and Manzoor A

    Momin Ahmad Khan, Novak Boskov, Fatima M. Anwar, and Manzoor A. Khan

  13. [13]

    InMechanistic Interpretability Workshop at NeurIPS 2025

    Controlling Vision–Language–Action Policies through Sparse Latent Directions. InMechanistic Interpretability Workshop at NeurIPS 2025. https: //openreview.net/forum?id=wtf3ww1EOL

    2025

  14. [14]

    Hyesu Lim, Jinho Choi, Jaegul Choo, and Steffen Schneider. 2025. Sparse autoencoders reveal selective remapping of visual concepts during adapta- tion. InThe Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=imT03YXlG2

    2025

  15. [15]

    Andre Martins and Ramon Astudillo. 2016. From Softmax to Sparsemax: A Sparse Model of Attention and Multi-Label Classification. InProceedings of The 33rd International Conference on Machine Learning (Proceedings of Machine Learn- ing Research, Vol. 48), Maria Florina Balcan and Kilian Q. Weinberger (Eds.). PMLR, New York, New York, USA, 1614–1623. https:/...

    2016

  16. [16]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po- Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick La...

    2024

  17. [17]

    Mateusz Pach, Shyamgopal Karthik, Quentin Bouniot, Serge Belongie, and Zeynep Akata. 2025. Sparse Autoencoders Learn Monosemantic Features in Vision- Language Models. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems. https://openreview.net/forum?id=DaNnkQJSQf

    2025

  18. [18]

    Ben Peters, Vlad Niculae, and André F. T. Martins. 2019. Sparse Sequence-to- Sequence Models. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Anna Korhonen, David Traum, and Lluís Màrquez (Eds.). Association for Computational Linguistics, Florence, Italy, 1504–1519. doi:10.18653/v1/P19-1146

    doi:10.18653/v1/p19-1146 2019

  19. [19]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs.CV] https://arxiv.org/ abs/2103.00020

    Pith/arXiv arXiv 2021

  20. [20]

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda. 2024. Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders. arXiv:2407.14435 [cs.LG] https://arxiv.org/abs/2407.14435

    Pith/arXiv arXiv 2024

  21. [21]

    Samuel Stevens, Wei-Lun Chao, Tanya Berger-Wolf, and Yu Su. 2025. Interpretable and Testable Vision Features via Sparse Autoencoders. arXiv:2502.06755 [cs.CV] https://arxiv.org/abs/2502.06755

    arXiv 2025

  22. [22]

    Zachary Teed and Jia Deng. 2020. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. arXiv:2003.12039 [cs.CV] https://arxiv.org/abs/2003.12039

    Pith/arXiv arXiv 2020

  23. [23]

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, et al

  24. [24]

    Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet.Transformer Circuits Thread(2024)

    2024

  25. [25]

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. 2022. VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training. arXiv:2203.12602 [cs.CV] https://arxiv.org/abs/2203.12602

    Pith/arXiv arXiv 2022

  26. [26]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2019. Representation Learning with Contrastive Predictive Coding. arXiv:1807.03748 [cs.LG] https://arxiv.org/ abs/1807.03748

    Pith/arXiv arXiv 2019

  27. [27]

    Zerui Wang and Yan Liu. 2024. STAA: Spatio-Temporal Attention Attribution for Real-Time Interpreting Transformer-based Video Models. arXiv:2411.00630 [cs.CV] https://arxiv.org/abs/2411.00630

    Pith/arXiv arXiv 2024

  28. [28]

    Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and XIAOJUAN QI. 2022. Self-Supervised Visual Representation Learning with Semantic Grouping. In Advances in Neural Information Processing Systems, Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (Eds.). https://openreview.net/forum? id=H3JObxjd8S

    2022

  29. [29]

    Yujia Zhang, Lai-Man Po, Xuyuan Xu, Mengyang Liu, Yexin Wang, Weifeng Ou, Yuzhi Zhao, and Wing-Yin Yu. 2022. Contrastive Spatio-Temporal Pretext Learning for Self-Supervised Video Representation. InAAAI

    2022