Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Low-latency Event-based Object Detection with Spatially-Sparse Linear Attention

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read SSLA-Det claims spatially-sparse linear attention makes asynchronous event-based detection accurate and cheap, cutting per-event compute more than 20x while setting state-of-the-art results on Gen1 and N-Caltech101.

desk verdict SSLA is a real architectural contribution with a sound parallel-training trick, but the Gen1 evaluation reporting is sloppy enough that I'd want the internal numbers cleaned up before trusting the exact mAP claims. read the letter →

arxiv 2603.06228 v3 pith:DROH2DSQ submitted 2026-03-06 cs.CV

Haiqing Hao , Zhipeng Sui , Rong Zou , Zijia Dai , Nikola Zubić , Davide Scaramuzza , Wenhui Wang This is my paper
classification cs.CV
keywords eventcamerasasynchronousinferencelinearattentionobjectdetectionspatialsparsitymixture-of-spacesscatter-compute-gatherlowlatency
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

Event cameras emit a sparse, time-stamped stream of pixel events; a detector that updates on every event can react faster, but existing asynchronous models are either hard to train on long sequences or become computationally expensive as they get accurate. This paper tries to remove that trade-off with Spatially-Sparse Linear Attention (SSLA): instead of one global state that every event updates, SSLA keeps many small patch states and lets each event touch only the patches covering its location. A position-aware projection feeds the event's relative position within each patch into the model, and a scatter-compute-gather procedure reorganizes events into patch-specific subsequences so training stays parallel while inference remains event-by-event. The authors build SSLA-Det from this module and report state-of-the-art asynchronous detection accuracy on Gen1 (0.375 mAP) and N-Caltech101 (0.515 mAP), with more than a 20x reduction in per-event computation versus the strongest prior asynchronous detector. If the parallel training procedure truly matches recurrent inference, this makes low-latency event-based object detection both accurate and efficient.

What carries the argument

Linear attention, used throughout, means a parallel-trainable linear recurrent model whose state update is a linear combination of the previous state and a projection of the current input; this dual parallel/recurrent form is what allows one set of weights to train on long sequences and run event-by-event. The carrying mechanism is the mixture-of-spaces (MOS) state decomposition: the spatial domain is divided into overlapping local patches, each holding an independent attention state, and each event scatters into only the few patches that cover its coordinates, computed in parallel, then gathers by summation. The position-aware projection (PAP) indexes a learnable linear map by the event's r

What would settle it

Take a trained SSLA-Det model and run the same event sequence twice: once through Algorithm 1's parallelized training path and once through the event-by-event recurrent path with the same weights, then compare the patch states and final detections; any mismatch falsifies the equivalence that underpins the training-efficiency claim. A second, independent check is to reproduce the reported Gen1 mAP of 0.375 under the paper's data split and training configuration, since the paper reports no such reproducibility test.

Watch

Extended reading notes

Core claim

The central claim is that state-level spatial sparsity makes linear attention viable for event-based object detection without sacrificing parallel training or recurrent inference. SSLA decomposes the spatial domain into overlapping P by P patches, each with its own linear-attention state; an incoming event activates exactly the patches that contain its coordinates (a constant number after padding), updates those states, and sums their outputs, so computed work per event does not grow with image resolution. The paper reports that replacing this with a standard linear attention, at the same embedding width or at similar FLOPs, drops mAP to near zero on Gen1, because the fine-grained spatial st

Load-bearing premise

The load-bearing premise is that the scatter-compute-gather training procedure (Algorithm 1) produces outputs identical to online event-by-event recurrent inference; the paper supplies no proof and no empirical comparison of the two modes, and if the equivalence fails, the reported parallel-training efficiency would not transfer to the low-latency inference the paper claims.

Editorial extensions

If this is right

  • Event-by-event object detection can be both accurate and cheap: SSLA-Det reports the best asynchronous mAP on Gen1 and N-Caltech101 while using more than 20x fewer FLOPs per event than the previous best asynchronous detector.
  • The entire pipeline, backbone and detection head, is asynchronous; predictions update only at the arriving event's location, and per-event work is independent of image resolution.
  • Spatial sparsity, not model size, is doing the work: the ablation shows standard linear attention at similar FLOPs drops detection performance to near zero on Gen1, while SSLA keeps a fine-grained state hundreds of times larger for the same compute.
  • The position-aware projection is necessary for spatial reasoning: removing both input and output projections drops Gen1 mAP from about 0.335 to 0.014.
  • Patch size is a direct accuracy-latency dial: increasing from P=2 to P=4 raises mAP from 0.200 to 0.371 while increasing per-event FLOPs and training memory.

Reading between the lines

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

  • The paper does not report a direct comparison between the scatter-compute-gather training path and the event-by-event recurrent path on the same weights; testing that equivalence would settle whether the claimed parallel-training advantage truly transfers to low-latency deployment.
  • Because SSLA's per-event FLOPs do not depend on image resolution, the module should in principle extend to megapixel event sensors without raising per-event compute; the paper only demonstrates 240x180 and 304x240 inputs, leaving megapixel scaling untested.
  • The authors note SSLA could be combined with frame-based vision by injecting image features into intermediate layers; that hybrid direction is not evaluated, but it is a natural next test if the sparse linear-attention states remain stable under dense feature injection.
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 introduces Spatially-Sparse Linear Attention (SSLA), a linear-attention module that decomposes the global recurrent state into a bank of spatially localized sub-states via a mixture-of-spaces structure. Each event activates only the patches containing its coordinates, and a position-aware projection encodes the event's relative position within each patch. To retain parallel training, the authors derive a scatter-compute-gather procedure that stably reorders event embeddings into patch-specific subsequences, applies independent linear attention per patch, and gathers the interim outputs. On this module they build SSLA-Det, an end-to-end asynchronous detector with a four-stage backbone and a modified YOLOX head, and report state-of-the-art asynchronous mAP on Gen1 (0.375) and N-Caltech101 (0.515), with per-event FLOPs reduced by more than 20x relative to the strongest prior asynchronous baseline DAGr-L. Ablations study spatial sparsity, position-aware projection, patch size, training speed, and inference latency, and a limitation section acknowledges the remaining accuracy gap to synchronous methods and the memory bottleneck for scaling.

Significance. If the numbers hold, this is a significant contribution: it appears to be the first end-to-end asynchronous linear-attention detector for event cameras, and it directly addresses the parallel-recurrent bottleneck and the accuracy-efficiency trade-off that have limited prior asynchronous models. The core design is clearly presented, the mixture-of-spaces idea is well motivated by event sparsity, and the ablations support the importance of both the sparse state decomposition and the position-aware projection. The reported training-speedup and inference-latency measurements are also valuable. However, the evaluation reporting contains internal inconsistencies that currently prevent verification of the central empirical claims, so the contribution cannot be fully assessed in its present form.

major comments (4)
  1. [§4.2, Tables 5 and 6] SSLA-S with full PAP is reported as 0.355 mAP in Table 5 and as 0.335 mAP in Table 6, while AP50 (0.610) and AP75 (0.322) are identical in both tables. For the same model and metric, identical AP50/AP75 with different mAP is impossible, indicating a metric-calculation error or a misreported value. Since Table 1 also lists SSLA-S at 0.334 mAP without stating the split, the reliability of all mAP values — including the headline 0.375 for SSLA-L — is in question. Please correct the tables, state the evaluation split for Table 1, and add error bars or multiple-seed results.
  2. [§4.1, Table 1] The Gen1 evaluation protocol is not fully specified. The paper never states whether the results in Table 1 are on the training, validation, or test split, nor whether the baseline numbers (DAGr-N/S/M/L, etc.) were obtained on the same split. Table 5 explicitly reports 'validation set of Gen1' and gives SSLA-S mAP=0.355, while Table 1 gives SSLA-S mAP=0.334. If Table 1 is test, this must be stated and compared with published test numbers; if it is validation, the discrepancy must be reconciled. The claimed SOTA advantage over DAGr-L rests on this comparison.
  3. [§3.3.3, Algorithm 1] The scatter-compute-gather training procedure is asserted to reproduce online event-by-event recurrent inference, but no formal proof or empirical verification is provided. The claim is plausible because stable sorting by patch index preserves temporal order within each patch subsequence; however, this equivalence is load-bearing for the parallel-training advantage and for the transfer of trained models to the low-latency recurrent setting. Please add an explicit proof sketch or a numerical check comparing Algorithm 1 outputs with a direct recurrent implementation on a small sequence.
  4. [§3.4, Fig. 3] The asynchronous YOLOX head is described only as changing the stem convolution to 1x1 and updating predictions at the event position. A standard YOLOX head contains 3x3 convolutions that require spatial context; it is not clear how the entire head is made asynchronous without either restricting all layers to 1x1 convolutions or introducing additional spatial dependencies. Please specify the exact head architecture or cite a reference that details it.
minor comments (5)
  1. [§3.3.3, Algorithm 1] Step 7 uses stable sorting by patch index. It would help to explicitly state that stable sorting preserves the temporal order of events within each patch subsequence, since this is the key to the claimed equivalence with recurrent inference.
  2. [§4.3, Table 3] The evaluation split and random seeds for the LSTM comparison are not given. SSLA-S mAP differs among Table 1 (0.334), Table 3 (0.335), and Table 5 (0.355); clarifying the protocol would remove ambiguity.
  3. [Eq. (5)] The notation W_in[δ_{i,k}] could be clearer; specifying that δ indexes a tensor slice of W_in would help the reader understand the parameter count and the spatial-prior mechanism.
  4. [§5] The limitation section is appreciated. Consider adding a comment on how the PAP parameter count scales with P^2 and whether this constrains the choice of patch size at scale.
  5. [Figure 1] The labels in the scatter/gather diagram are small; enlarging them or using a legend would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: SSLA-Det's claims are empirical, self-citations are contextual baselines, and the sparse-attention derivation does not reduce to its inputs.

full rationale

The paper's central accuracy/efficiency claims are empirical evaluations against published baselines on Gen1 and N-Caltech101; no mAP is used as a fitting target or derived from the SSLA equations. The scatter-compute-gather procedure (Sec. 3.3.3, Algorithm 1) is an implementation identity for independent per-patch states: because patches are spatially disjoint in state and subsequences preserve temporal order via stable sorting, the gathered outputs are constructed to match the recurrent event-by-event update. This is a claimed equivalence, not a circular definition, though it is not formally proved—an omitted verification, not evidence of circularity. Self-citations ([14] EVA, [11] DAGr, [25] AsyNet, [37] AEGNN, [49] ERGO-12, [41] S7) are used to supply baselines or related-work context; the baseline numbers come from externally published papers and the SSLA design explicitly credits [6] Mixture-of-Memories and [36] FARSE-CNN rather than importing its architecture from a self-citation. Section 5 honestly acknowledges limitations (gaps to synchronous methods, no large-scale scaling study); these scope statements do not conceal a reduction of the result to its inputs. The undefined Gen1 split and the mAP/AP50/AP75 inconsistency in Tables 5–6 are evaluation-protocol/correctness risks, outside the circularity definition: if the numbers are wrong, the failure is experimental, not a derivation equivalent to its own assumptions. No step meets the 'quote and exhibit reduction' bar, so the score is 0.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the design of spatially decomposed states, the algorithmic equivalence of scatter-compute-gather with online inference, and the validity of using local patches to capture spatial context for detection. All are validated only empirically on two datasets; no theoretical guarantees or external confirmations are provided.

free parameters (4)
  • Patch size P = 3 (default)
    Chosen by validation; P=2 gives mAP 0.200 and P=4 gives 0.371 (Tab. 7). Central to the accuracy-efficiency trade-off.
  • First-stage embedding dimension D_out = 12 (S), 16 (B), 24 (M), 32 (L)
    Manually scaled to create model variants; not selected by a principled rule.
  • Backbone stages and SSLA layers per stage = 4 stages, 2 layers each
    Fixed architecture choice, not ablated in the paper.
  • Temporal dropout keep ratio = rho ~ U(0.8,1.0)
    Stochastic augmentation used during training; affects effective sequence length and is borrowed from FARSE-CNN.
assumptions (6)
  • standard math Causal linear attention has an equivalent parallel-scan and recurrent form (Eq. 1).
    Used as the backbone update rule; established by prior work [13,20,46], not proven in this paper.
  • domain assumption Event data is a temporally ordered sequence e_i=(x_i,t_i,p_i).
    Standard event-camera representation; stated in Section 3.1.
  • ad hoc to paper Sliding-window patches of size P with stride 1 and padding give each event exactly A=P^2 active patches.
    Core design of the mixture-of-spaces structure; ensures constant per-event compute, but is not justified theoretically.
  • ad hoc to paper Independent per-patch states that share parameters and are aggregated by summation form a valid causal spatiotemporal model.
    Not proven; supported only by empirical ablation (Tab. 5).
  • ad hoc to paper Scatter-compute-gather (Algorithm 1) yields outputs identical to online event-by-event inference.
    Assumed equivalence; not formally proven. If false, training and inference would be inconsistent.
  • domain assumption Temporal dropout and sparse pooling from FARSE-CNN can be inserted asynchronously without breaking causal event-by-event updates.
    Borrowed from cited work; no ablation of these components in this paper.
invented entities (1)
  • Mixture-of-Spaces (MOS) patch-state bank
    purpose: Replaces a single global linear-attention state with one independent state per spatial patch, enabling P^2-active sparse updates.
    Architectural invention whose validity is demonstrated only on the in-paper benchmarks; no external falsifiable prediction outside the paper's datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-latency Event-based Object Detection with Spatially-Sparse Linear Attention." pith.science (2026). https://pith.science/paper/DROH2DSQ

@misc{pith2026260306228,
  author       = {Pith},
  title        = {Pith review of: Low-latency Event-based Object Detection with Spatially-Sparse Linear Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DROH2DSQ}},
  note         = {Machine review of arXiv:2603.06228}
}
read the original abstract

Event cameras provide sequential visual data with spatial sparsity and high temporal resolution, making them attractive for low-latency object detection. Existing asynchronous event-based neural networks exploit this low-latency advantage by updating predictions event by event, but still suffer from two bottlenecks: recurrent architectures are difficult to train efficiently on long sequences, and improving accuracy often increases per-event computation and latency. Linear attention is appealing because it enables parallel training and recurrent inference. However, its dense state updates make per-event computation scale with the state size, yielding a poor accuracy-efficiency trade-off for object detection, where accurate localization requires fine-grained spatial states. The key challenge is therefore to introduce sparse state activation that exploits the spatial sparsity of events while preserving efficient parallel training. We propose Spatially-Sparse Linear Attention (SSLA), which introduces a mixture-of-spaces state decomposition and a scatter-compute-gather training procedure, enabling state-level sparsity as well as training parallelism. Building on SSLA, we develop an end-to-end asynchronous linear attention model, SSLA-Det, for low-latency event-based object detection. On Gen1 and N-Caltech101, SSLA-Det achieves state-of-the-art accuracy among asynchronous methods, reaching 0.375 mAP and 0.515 mAP, respectively, while reducing per-event computation by over 20 times compared with the strongest prior asynchronous baseline, demonstrating the potential of linear attention for low-latency event-based vision.

Figures

Figures reproduced from arXiv: 2603.06228 by the authors.

Figure 1
Figure 1. Our method processes (a) asynchronous event sequence with (b) a sparsely activated linear attention neural network for (c) low [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Spatially-Sparse Linear Attention (SSLA) Module. (1) and (2) are 2 × 2 sliding window patch examples on the spatial domain. Each event is (a) scattered into multiple patches that spatially cover it (e.g. ei to patch (1) and (2). For brevity, we did not show other patches that cover ei). Each patch gets a subsequence of the events that it covers, and computes the interim outputs independently (b) with… view at source ↗
Figure 3
Figure 3. Overview of the SSLA-Det model. Left: Fully asynchronous event-based detector. The input data is processed by a 4-stage asynchronous backbone. Each stage contains 2 SSLA layers followed by sparse pooling and temporal dropout. Right: SSLA layer. Input event embeddings are processed by the SSLA module, including two position-aware projections, scatter, parallel patch-wise linear attention and gather. A residual connec… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the detection results on the Gen1 dataset. Green boxes denote ground truth and orange boxes denote predictions with confidence scores. Predicted boxes with confidence scores below 0.5 are removed. (a)-(d): Cars. (e)-(h): Pedestrians. (i)-(l): Failure c…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 7 linked inside Pith

  1. [1]

    Layer normalization.arXiv preprint arXiv:1607.06450,

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hin- ton. Layer normalization.arXiv preprint arXiv:1607.06450,

    arXiv

  2. [2]

    Ehgcn: Hierarchical euclidean-hyperbolic fusion via motion-aware gcn for hybrid event stream perception.arXiv preprint arXiv:2504.16616, 2025

    Haosheng Chen, Lian Luo, Mengjingcheng Mo, Zhanjie Wu, Guobao Xiao, Ji Gan, Jiaxu Leng, and Xinbo Gao. Ehgcn: Hierarchical euclidean-hyperbolic fusion via motion-aware gcn for hybrid event stream perception.arXiv preprint arXiv:2504.16616, 2025. 2, 8

    arXiv 2025

  3. [3]

    Hugnet: Hemi- spherical update graph neural network applied to low-latency event-based optical flow

    Thomas Dalgaty, Thomas Mesquida, Damien Joubert, Amos Sironi, Pascal Vivet, and Christoph Posch. Hugnet: Hemi- spherical update graph neural network applied to low-latency event-based optical flow. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3953–3962, 2023. 2

    2023

  4. [4]

    Graph neural network combining event stream and periodic aggre- gation for low-latency event-based vision

    Manon Dampfhoffer, Thomas Mesquida, Damien Joubert, Thomas Dalgaty, Pascal Vivet, and Christoph Posch. Graph neural network combining event stream and periodic aggre- gation for low-latency event-based vision. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6909–6918, 2025. 2

    2025

  5. [5]

    A large scale event- based detection dataset for automotive.arXiv preprint arXiv:2001.08499, 2020

    Pierre De Tournemire, Davide Nitti, Etienne Perot, Da- vide Migliore, and Amos Sironi. A large scale event- based detection dataset for automotive.arXiv preprint arXiv:2001.08499, 2020. 5

    arXiv 2001

  6. [6]

    Mom: Linear sequence modeling with mixture- of-memories

    Jusen Du, Weigao Sun, Disen Lan, Jiaxi Hu, Tao Zhang, and Yu Cheng. Mom: Linear sequence modeling with mixture- of-memories. InThe Fourteenth International Conference on Learning Representations, 2026. 2, 3

    2026

  7. [7]

    Dy- namic obstacle avoidance for quadrotors with event cameras

    Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dy- namic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1

    2020

  8. [8]

    Eventpillars: Pillar-based efficient representations for event data

    Rui Fan, Weidong Hao, Juntao Guan, Lai Rui, Lin Gu, Tong Wu, Fanhong Zeng, and Zhangming Zhu. Eventpillars: Pillar-based efficient representations for event data. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 2861–2869, 2025. 6, 8, 9

    2025

Show all 49 references
  1. [9]

    Event-based vision: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(1):154–180, 2020

    Guillermo Gallego, Tobi Delbr ¨uck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, J ¨org Conradt, Kostas Daniilidis, et al. Event-based vision: A survey.IEEE transactions on pattern analysis and machine intelligence, 44(1):...

    2020

  2. [10]

    Yolox: Exceeding yolo series in 2021.arXiv preprint arXiv:2107.08430, 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021.arXiv preprint arXiv:2107.08430, 2021. 5

    2021 arXiv

  3. [11]

    Low-latency auto- motive vision with event cameras.Nature, 629(8014):1034– 1040, 2024

    Daniel Gehrig and Davide Scaramuzza. Low-latency auto- motive vision with event cameras.Nature, 629(8014):1034– 1040, 2024. 1, 2, 5, 6, 7, 8

    2024

  4. [12]

    Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012

    Alex Graves. Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012. 6

    2012

  5. [13]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. InFirst conference on lan- guage modeling, 2024. 1, 3

    2024

  6. [14]

    Maximizing asyn- chronicity in event-based neural networks

    Haiqing Hao, Nikola Zubic, Weihua He, Zhipeng Sui, Da- vide Scaramuzza, and Wenhui Wang. Maximizing asyn- chronicity in event-based neural networks. InThe Four- teenth International Conference on Learning Representa- tions, 2026. 1, 2, 3, 6, 8

    2026

  7. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

    2016

  8. [16]

    Neuromorphic-enabled video- activated cell sorting.Nature communications, 15(1):10792,

    Weihua He, Junwen Zhu, Yongxiang Feng, Fei Liang, Kaichao You, Huichao Chai, Zhipeng Sui, Haiqing Hao, Guoqi Li, Jingjing Zhao, et al. Neuromorphic-enabled video- activated cell sorting.Nature communications, 15(1):10792,

  9. [17]

    Averaging weights 9 leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights 9 leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018. 6

    2018 arXiv

  10. [18]

    Kamil Jeziorek, Piotr Wzorek, Krzysztof Blachut, Hiroshi Nakano, Manon Dampfhoffer, Thomas Mesquida, Hiroaki Nishi, Thomas Dalgaty, and Tomasz Kryjak. Hardware- accelerated graph neural networks: an alternative approach for neuromorphic event-based audio classification and key...

  11. [19]

    As- sociative memory augmented asynchronous spatiotemporal representation learning for event-based perception

    Uday Kamal, Saurabh Dash, and Saibal Mukhopadhyay. As- sociative memory augmented asynchronous spatiotemporal representation learning for event-based perception. InThe Eleventh International Conference on Learning Representa- tions, 2023. 2, 3

    2023

  12. [20]

    Transformers are rnns: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc ¸ois Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational confer- ence on machine learning, pages 5156–5165. PMLR, 2020. 1, 3

    2020

  13. [21]

    Asynchronous collaborative graph representation for frames and events

    Dianze Li, Jianing Li, Xu Liu, Xiaopeng Fan, and Yonghong Tian. Asynchronous collaborative graph representation for frames and events. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 1655–1666,

  14. [22]

    Graph-based asyn- chronous event processing for rapid object recognition

    Yijin Li, Han Zhou, Bangbang Yang, Ye Zhang, Zhaopeng Cui, Hujun Bao, and Guofeng Zhang. Graph-based asyn- chronous event processing for rapid object recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 934–943, 2021. 2, 6, 8

    2021

  15. [23]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 6

    2014

  16. [24]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2019. 5

    2019

  17. [25]

    Event-based asynchronous sparse con- volutional networks

    Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse con- volutional networks. InEuropean Conference on Computer Vision, pages 415–431. Springer, 2020. 2, 6, 8

    2020

  18. [26]

    Converting static image datasets to spiking neuromorphic datasets using saccades.Frontiers in neuro- science, 9:437, 2015

    Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades.Frontiers in neuro- science, 9:437, 2015. 5

    2015

  19. [27]

    Resurrecting recurrent neural networks for long se- quences

    Antonio Orvieto, Samuel L Smith, Albert Gu, Anushan Fernando, Caglar Gulcehre, Razvan Pascanu, and Soham De. Resurrecting recurrent neural networks for long se- quences. InInternational conference on machine learning, pages 26670–26698. PMLR, 2023. 5

    2023

  20. [28]

    Scaling linear attention with sparse state expansion

    Yuqi Pan, Yongqi An, Zheng Li, Yuhong Chou, Rui-Jie Zhu, Xiaohui Wang, Mingxuan Wang, Jinqiao Wang, and Guoqi Li. Scaling linear attention with sparse state expansion. In The Fourteenth International Conference on Learning Rep- resentations, 2026. 3

    2026

  21. [29]

    Fully neuromorphic vision and control for au- tonomous drone flight.Science Robotics, 9(90):eadi0591,

    Federico Paredes-Vall ´es, Jesse J Hagenaars, Julien Dupey- roux, Stein Stroobants, Yingfu Xu, and Guido CHE de Croon. Fully neuromorphic vision and control for au- tonomous drone flight.Science Robotics, 9(90):eadi0591,

  22. [30]

    Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, ...

    2019

  23. [31]

    Rwkv: Reinventing rnns for the transformer era

    Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Leon Derczynski, et al. Rwkv: Reinventing rnns for the transformer era. InFindings of the association for computational linguistics: EMNLP 2023, pages...

    2023

  24. [32]

    Get: Group event transformer for event-based vision

    Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: Group event transformer for event-based vision. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 6038–6048, 2023. 1, 6, 8

    2023

  25. [33]

    Scene adaptive sparse transformer for event-based object detection

    Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 16794–16804, 2024. 6, 8

    2024

  26. [34]

    Learning to detect objects with a 1 megapixel event camera.Advances in Neural Information Processing Systems, 33:16639–16652, 2020

    Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera.Advances in Neural Information Processing Systems, 33:16639–16652, 2020. 5

    2020

  27. [35]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  28. [36]

    Farse-cnn: Fully asynchronous, recurrent and sparse event-based cnn

    Riccardo Santambrogio, Marco Cannici, and Matteo Mat- teucci. Farse-cnn: Fully asynchronous, recurrent and sparse event-based cnn. InEuropean conference on computer vi- sion, pages 1–18. Springer, 2024. 1, 2, 5, 6, 8, 9

    2024

  29. [37]

    Aegnn: Asynchronous event-based graph neural networks

    Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 12371–12381, 2022. 1, 2, 6, 8

    2022

  30. [38]

    Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models

    Mark Sch ¨one, Neeraj Mohan Sushma, Jingyue Zhuge, Chris- tian Mayr, Anand Subramoney, and David Kappel. Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models. In2024 International Con- ference on Neuromorphic Systems (ICONS), pages 12...

    2024

  31. [39]

    Eventnet: Asynchronous recursive event processing

    Yusuke Sekikawa, Kosuke Hara, and Hideo Saito. Eventnet: Asynchronous recursive event processing. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3887–3896, 2019. 1, 2

    2019

  32. [40]

    Col2a: Convolution-free local linear attention for spatiotemporal event processing

    Yusuke Sekikawa, Jun Nagata, Itsumi Araki, and Andreu Girbau. Col2a: Convolution-free local linear attention for spatiotemporal event processing. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4869–4880, 2026. 3 10

    2026

  33. [41]

    S7: Selective and simplified state space layers for sequence modeling.arXiv preprint arXiv:2410.03464, 2024

    Taylan Soydan, Nikola Zubi ´c, Nico Messikommer, Sid- dhartha Mishra, and Davide Scaramuzza. S7: Selective and simplified state space layers for sequence modeling.arXiv preprint arXiv:2410.03464, 2024. 1, 2, 3

    2024 arXiv

  34. [42]

    Tri- ton: an intermediate language and compiler for tiled neu- ral network computations

    Philippe Tillet, Hsiang-Tsung Kung, and David Cox. Tri- ton: an intermediate language and compiler for tiled neu- ral network computations. InProceedings of the 3rd ACM SIGPLAN International Workshop on Machine Learning and Programming Languages, pages 10–19, 2019. 5

    2019

  35. [43]

    ALERT- transformer: Bridging asynchronous and synchronous ma- chine learning for real-time event-based spatio-temporal data

    Carmen Martin Turrero, Maxence Bouvier, Manuel Bre- itenstein, Pietro Zanuttigh, and Vincent Parret. ALERT- transformer: Bridging asynchronous and synchronous ma- chine learning for real-time event-based spatio-temporal data. InForty-first International Conference on Machine L...

    2024

  36. [44]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 3

    2017

  37. [45]

    Smamba: Sparse mamba for event- based object detection

    Nan Yang, Yang Wang, Zhanwen Liu, Meng Li, Yisheng An, and Xiangmo Zhao. Smamba: Sparse mamba for event- based object detection. InProceedings of the AAAI Confer- ence on Artificial Intelligence, pages 9229–9237, 2025. 6, 8

    2025

  38. [46]

    Gated linear attention trans- formers with hardware-efficient training.arXiv preprint arXiv:2312.06635, 2023

    Songlin Yang, Bailin Wang, Yikang Shen, Rameswar Panda, and Yoon Kim. Gated linear attention trans- formers with hardware-efficient training.arXiv preprint arXiv:2312.06635, 2023. 1, 3

    2023 arXiv

  39. [47]

    Parallelizing linear transformers with the delta rule over sequence length.Advances in neural information processing systems, 37:115491–115522, 2024

    Songlin Yang, Bailin Wang, Yu Zhang, Yikang Shen, and Yoon Kim. Parallelizing linear transformers with the delta rule over sequence length.Advances in neural information processing systems, 37:115491–115522, 2024. 1

    2024

  40. [48]

    Compute-in- memory implementation of state space models for event se- quence processing.Nature Communications, 2026

    Xiaoyu Zhang, Mingtao Hu, Sen Lu, Soohyeon Kim, Eric Yeu-Jer Lee, Yuyang Liu, and Wei D Lu. Compute-in- memory implementation of state space models for event se- quence processing.Nature Communications, 2026. 7

    2026

  41. [49]

    From chaos comes order: Ordering event rep- resentations for object recognition and detection

    Nikola Zubi ´c, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event rep- resentations for object recognition and detection. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 12846–12856, 2023. 1, 6, 8 11

    2023

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.