{"id":"99d4f22f-7748-4a96-a562-52679b103e9b","arxiv_id":"2508.19806","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A learned input-adaptive threshold gates neuron activations in event-based neural networks, delivering near-SOTA object detection and optical flow with 32-68% fewer synaptic operations.","lead":"This paper introduces a way to make event-camera neural networks smarter about which neurons to fire, based on the input scene. The result is a model that matches or beats leading methods on object detection and optical flow while using far fewer operations, an efficiency win for robots and edge devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Recurrent threshold in Eq. 2b ignores current input, risking suppression of newly appearing features; empirical wins on two benchmarks don't establish when masking fails.","rationale":"The paper's strongest evidence is the head-to-head comparison of CSSL-SEED-256 with SEED-256 (both with GRU and MGU variants) and the ablation in Table IV showing CSSL outperforms a ReLU-based counterpart with lower GSOp. These are real, independent supports. However, the central mechanism is the thresholding operation, and its design in the recurrent branch is the least secure part. The threshold in Eq. 2b ignores the current input; this is not a minor implementation detail but a structural limitation that could produce systematic blindness to newly appearing features. The two tasks in the paper are automotive detection and optical flow on drone/outdoor scenes; both have specific motion statistics, and the model may have tuned b_v to a compromise that works there. The paper provides no analysis of failure modes, so we cannot tell whether the method generalizes. The proposed check uses the released code to measure how often novel-input-driven activations are suppressed at previously inactive locations; if the fraction is negligible or uncorrelated with misses, the concern is refuted. This does not change the verdict: the reported results are credible, but the mechanism's generality remains conditional.","tokens_in":10514,"tokens_out":11934,"duration_ms":120804,"concrete_test":"Using the released object-detection code, instrument the trained CSSL-SEED-256 model to log, on the 1Mpx test set, the recurrent-layer threshold v_th(t) and pre-threshold activation \\tilde y(t) for all timesteps. Filter to locations where the previous hidden state y(t-1) is zero and the current input x(t) is nonzero (newly active region). Measure the proportion of these locations for which H(\\tilde y - v_th)=0 (i.e., the current input is killed). Then overlay these killed locations with ground-truth bounding boxes. If a substantial fraction of true-object pixels are killed at the timestep the object first appears, the concern that the recurrent threshold suppresses novel informative features is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the assumption that the learned threshold v_th reliably separates informative activations from noise in every layer and timestep. For recurrent modules (Eq. 2b), v_th(t)=σ(W_v y(t-1)+b_v) depends only on the previous sparse hidden state, not on the current input x(t). At a pixel that was inactive at t-1 (e.g., an object entering the field of view or an edge moving into a previously empty region), v_th(t) reduces to σ(b_v), a learned constant. If that constant is large, all current-input-driven activations below it are zeroed by s(t)=H(c(t)-v_th(t)); because y(t)=c(t)⊙s(t) is the only signal passed onward, the network is blind to the new feature. The paper reports aggregate mAP/AEE on two benchmarks but offers no analysis of how often this destructive masking occurs or whether such cases correlate with misses. Since the method's name and abstract claim thresholds are based on the input distribution, this blind spot in the recurrence is a genuine gap: the empirical wins could be specific to the benchmark's motion statistics rather than a general principle.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Context-aware Sparse Spatiotemporal Learning (CSSL), a training/inference mechanism that adds per-pixel, input-dependent threshold convolutions to standard convolutional and convolutional-recurrent units. The threshold is produced by a sigmoid-bounded convolution (Eq. 1) and applied through a Heaviside gate, giving sparse activations without an explicit sparsity loss. The method is instantiated in two architectures: CSSL-SEED for event-based object detection and CSSL-EV-FlowNet for optical flow. On the 1Mpx and Gen1 detection benchmarks it reports mAP close to or above the SEED baseline at lower GSOp, and on MVSEC it reports lower AEE and lower activation density than RNN-EV-FlowNet. The authors argue that CSSL removes the need to tune sparsity-loss weights and is suitable for neuromorphic deployment.","tokens_in":10836,"tokens_out":6541,"duration_ms":75609,"significance":"If the reported results hold, CSSL is a useful and simple technique for obtaining high activation sparsity in event-based vision networks, and the code release supports reproducibility. The comparison against an explicit sparsity-loss baseline (Table IV) is informative and shows that the learned thresholds attain a better accuracy/efficiency trade-off than L1 sparsification. However, the strength of the evidence is limited by single-run experiments with small margins, by the unclear cost accounting of the auxiliary threshold convolutions, and by the fact that the recurrent threshold in Eq. (2b) is not actually conditioned on the current input. The core idea is plausible and the empirical results are encouraging, but the load-bearing efficiency and generality claims need stronger support before publication.","major_comments":[{"comment":"All reported detection and flow results appear to come from a single training run. The central claim of 'comparable or superior performance' rests on small differences, e.g., CSSL-SEED-256 46.4 vs. SEED-256 44.9 mAP on 1Mpx, and 46.3 vs. 45.3 on Gen1. Without error bars, multiple seeds, or significance tests it is impossible to know whether these gaps are systematic or training noise. Please report mean and std over at least 3 runs, or, if that is computationally infeasible, provide a justification and at least repeat the most critical comparisons (e.g., SEED-256 vs. CSSL-SEED-256 and RNN-EV-FlowNet vs. CSSL-EV-FlowNet).","section":"§III-B, Tables II and III"},{"comment":"The recurrent threshold v_th(t) depends only on the previous sparse hidden state y(t-1), not on the current input x(t). For a pixel that was inactive at t-1, v_th(t) reduces to sigma(b_v), a learned constant. If this constant is large, current-input-driven activations in c(t) are zeroed by the Heaviside gate in Eq. (2e), so newly appearing features can be suppressed before they propagate. The paper offers no analysis of how often this destructive masking occurs or whether it correlates with misses on the benchmark tasks. Since the abstract claims thresholds are based on the input distribution, this is a gap. Please add an ablation where v_th also depends on x(t) (e.g., concatenating x(t) with y(t-1) before the threshold convolution) and/or analyze failure cases to show that newly appearing objects are not systematically thresholded out.","section":"§II-D, Eq. (2b)"},{"comment":"The reported GSOp numbers do not clearly state whether they include the extra threshold-generating convolution W_v in Eq. (1) and the threshold convolution W_v in Eq. (2b). These are additional synaptic operations, so the efficiency comparison depends on this accounting. If these auxiliary convolutions are excluded from GSOp, the CSSL efficiency advantage is overstated. In addition, the sentence in Section II-D claiming that 'the dense tensor has a negligible impact on the total number of synaptic operations' is not quantified; the multiplications Wxf*x(t) and Wxh*x(t) in Eq. (2) still contribute to the total. Please provide an explicit cost model or measured operation counts that include all auxiliary convolutions.","section":"§II-B/Eq. (1), Table II"},{"comment":"The context-aware residual block is described only in prose and the figure; no equations are given for how the second convolution is split into threshold and dense components, how the post-accumulation threshold is computed, or how the sparse output is formed. This is not only a reproducibility issue but also makes it hard to verify that the residual-block implementation in Table II matches the described mechanism. Please add the corresponding equations or pseudocode.","section":"§II-C, Fig. 2(b)"}],"minor_comments":[{"comment":"Typo: 'struggle to match of performance' should be 'struggle to match the performance'.","section":"Abstract"},{"comment":"The manuscript frequently uses 'we proposed' where 'we propose' is intended. Please correct.","section":"Throughout"},{"comment":"Table V lists 'SEED without sparsity loss' with mAP 45.0, while Table II reports SEED-256 with mAP 44.9. Please clarify whether these are the same model or different configurations.","section":"Table V vs Table II"},{"comment":"The claim that CSSL 'eliminates the need for manually tuned sensitive sparsity hyperparameters' is supported only for the L1 sparsity weight. The surrogate gradient scale and the sigmoid threshold range are also hyperparameters, but their sensitivity is not analyzed. A short sensitivity study would strengthen the claim.","section":"§III-D"},{"comment":"The soft-reset mechanism in Eq. (3) is introduced as beneficial for forgetting and training stability, but no ablation isolates its contribution. Please report a comparison with and without the soft reset.","section":"§II-D/§III-C"},{"comment":"The relationship between the optical-flow training setup and the MVSEC evaluation protocol is taken directly from [7]; please state explicitly whether the reported RNN-EV-FlowNet numbers are copied from [7] or re-run, and if re-run, whether the same code/config was used.","section":"§III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is appropriate for a computer-vision or neuromorphic-computing venue, and the central idea is worth publishing if the evidence is shored up. My main concerns are methodological: single-run results, ambiguous GSOp accounting, and the recurrent threshold's blindness to current input. Regarding the same-group baselines (SEED and RNN-EV-FlowNet), I do not see evidence of unfair comparison, but I would recommend requiring the authors to specify exactly where the baseline numbers come from and to provide evaluation scripts or baselines so that reviewers/readers can verify fairness. Given the high training cost, asking for 3 seeds of every model may be excessive; the editor could request at least repeated runs of the headline comparisons and an ablation addressing the recurrent-threshold concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is a clean, practical idea: learn a pixel-wise threshold that masks activations in conv and recurrent layers, driving sparsity without an explicit sparsity loss. It works. On 1Mpx object detection, CSSL-SEED-256 gets 46.4 mAP at 2.80 GSOp versus SEED's 44.9 at 3.83, and the optical flow numbers improve too. What's genuinely new is the specific combination: learned sigmoid thresholds, Heaviside masking with surrogate gradients, and a soft-reset auxiliary state in the recurrent unit. That distinguishes it from Deja Vu-style contextual sparsity and from the authors' own SEED baseline. The ablation with and without sparsity loss is informative: CSSL avoids the sensitivity of beta_sparse tuning and still beats ReLU+L1 on both accuracy and compute. The per-layer activation density table is a useful diagnostic. Detection code is released.\n\nSoft spots, in order of importance. First, every result is a single run with no error bars or significance tests. That is a real weakness for a paper whose central claim is \"comparable or superior performance.\" Second, the efficiency accounting is incomplete. It reports GSOp for the main path but does not include the cost of the extra threshold convolutions, and there is no hardware measurement to confirm the claimed energy benefits. Third, the recurrent threshold in Eq. 2b depends only on the previous hidden state y(t−1), not on the current input x(t). The stress-test concern lands: at a pixel that was inactive at t−1, the threshold reduces to sigma(bv), a learned constant. If that constant is high, newly arriving features are zeroed before they can propagate. The network can in principle learn bv low, and the two benchmark wins suggest the issue is not catastrophic, but the paper offers no analysis of how often this masking misses events. The abstract says thresholds are based on the input distribution; for the recurrent modules, that is not true. That gap should be acknowledged and ideally tested. Fourth, the strong baselines come from the authors' own group, but they are legitimate and are better than older methods, so I do not see that as a flaw.\n\nWho is this for? Researchers working on efficient event-based perception and neuromorphic deployment. It is a useful within-subfield contribution, not a field reshapper. The paper deserves a serious referee. My recommendation: send it out, but ask for error bars, a more complete compute cost breakdown, and a discussion or small experiment probing the recurrent threshold's behavior on new features appearing in previously empty regions.","headline":"Solid empirical method for sparsity in event-based vision; watch the recurrent threshold blind spot and missing error bars.","tokens_in":11245,"tokens_out":2088,"would_cite":true,"duration_ms":26165,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Context-aware thresholding sparsifies event-vision networks without sparsity-loss tuning.","keywords":["event-based vision","neuromorphic computing","activation sparsity","context-aware thresholding","object detection","optical flow","sparse recurrent networks","event cameras"],"falsifier":"Run a trained CSSL-SEED on a low-contrast, small-object subset of an event-detection benchmark and compare it with the same architecture with all learned thresholds forced to zero at inference. If the dense version's mAP is materially higher, the hard threshold is discarding task-critical features; if it is not, the masking is information-preserving on that distribution.","tokens_in":10431,"feed_emoji":"⚡","tokens_out":9815,"duration_ms":99235,"temperature":0.7,"pith_summary":"The paper sets out to show that event-camera networks can become highly sparse without explicit sparsity regularization, simply by learning where to place the activation threshold. It introduces Context-aware Sparse Spatiotemporal Learning (CSSL), which computes a per-pixel threshold from the input feature distribution and zeros out every activation below that threshold. On event-based object detection, CSSL-SEED-256 reaches 46.4 mAP on the 1Mpx dataset with 2.80 GSOp, while the SEED-256 baseline it builds on reaches 44.9 mAP with 3.83 GSOp. On optical flow, CSSL-EV-FlowNet reaches 2.38 average endpoint error at 10.61% neuron density, compared with 2.88 and 16.90% for the recurrent baseline. The practical significance is that robot perception with event cameras could run at a fraction of the compute, without fragile sparsity-loss hyperparameters.","feed_headline":"Adaptive thresholds cut event-vision compute, keep accuracy","feed_subtitle":"Learned per-pixel gates keep event-vision networks sparse and accurate, no sparsity-loss tuning needed.","key_machinery":"The central object is the context-aware thresholding gate: a per-pixel threshold v_th = sigma(W_v x + b_v) produced by a convolution over the input, followed by a Heaviside mask s = H(y~ - v_th). It does the sparsification work that sparsity loss terms normally do. In recurrent blocks the threshold is computed from the previous sparse hidden state y(t-1) rather than the current input, and a soft-reset subtracts the threshold after the unit fires; this regulates spatiotemporal sparsity while keeping the dense input convolution cheap.","core_discovery":"CSSL's central claim is that dynamic context-aware thresholding can replace both fixed ReLU thresholds and explicit sparsity regularization in event-based vision. At each time step a small convolution reads the input feature map and emits a per-pixel threshold v_th = sigma(W_v x + b_v) constrained to [0,1]; a Heaviside step then produces a binary mask that zeroes every output activation below that threshold. Because the threshold tracks the input distribution, the network concentrates computation on informative locations and naturally keeps activation density low. The authors show the mechanism generalizes across convolutional, residual, and recurrent blocks and across object detection and o","pith_inferences":["Because the recurrent threshold in Eq. (2b) sees only the previous hidden state and not the current input, abrupt scene changes that generate many fresh events while the hidden state is stale are an untested stress case; the paper does not analyze this failure mode.","The masking operator is a hard zero-out, so measuring how much task-relevant information actually sits below the learned thresholds (e.g., by ablating the mask at inference) would reveal where the efficiency-accuracy trade-off truly lives.","The same input-conditional gating idea could be applied to attention-based event backbones, where sparsity would prune tokens or channels rather than pixels, but the paper only demonstrates convolutional and recurrent modules."],"forward_implications":["Sparsity becomes an architectural property rather than a training objective, so practitioners can drop the beta_sparse hyperparameter and the two-stage tuning it requires.","CSSL modules can replace ordinary convolutions, residual blocks, and recurrent cells in existing architectures; the paper demonstrates this on detection and optical flow with consistent efficiency gains.","The framework works across multiple recurrent cells (MGU, GRU, MinimalRNN), with MGU and GRU nearly matching each other and MinimalRNN trading some accuracy for fewer parameters and operations.","On neuromorphic processors that charge per synaptic operation or per active neuron, the reduced activation density should translate directly into lower energy and faster inference, which the paper positions as the motivation for the framework."],"supporting_citations":[{"why":"Supplies the sparse convolutional recurrent detector architecture that CSSL builds on and the SEED baseline it must beat.","marker":"[8]"},{"why":"Provides the recurrent optical-flow architecture and the activation-sparsification baseline that CSSL-EV-FlowNet is compared against.","marker":"[7]"},{"why":"Supplies the surrogate-gradient scheme and event-recurrent training method used to backpropagate through the Heaviside mask.","marker":"[12]"},{"why":"Provides the 1Mpx event-camera detection dataset and the evaluation protocol used for the detection experiments.","marker":"[15]"},{"why":"Provides the Gen1 automotive event-detection dataset used as a second detection benchmark.","marker":"[16]"},{"why":"Provides the strong transformer-based event-detector baselines whose synaptic-operation counts CSSL is compared with.","marker":"[21]"},{"why":"Provides the Minimal Gated Unit that CSSL uses as its primary recurrent cell.","marker":"[13]"},{"why":"Motivates input-dependent contextual sparsity, the general idea CSSL adapts to event-based vision.","marker":"[10]"}],"fun_headline_variants":["Context-aware thresholds make event-vision nets naturally sparse","No sparsity-loss tuning: adaptive gates slim event-vision AI","Adaptive thresholds eliminate sparsity-loss tuning in event vision","Context-aware thresholds make event vision sparse without tuning","Dynamic thresholds replace sparsity losses in event-vision nets"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a learned per-pixel threshold, hard-zeroing every activation below it, reliably separates informative from redundant signals in every layer and every task, and that in recurrent layers the threshold can be read from the previous hidden state alone.","fun_headline_variants_meta":{"raw":{"variants":["Context-aware thresholds make event-vision nets naturally sparse","No sparsity-loss tuning: adaptive gates slim event-vision AI","Adaptive thresholds eliminate sparsity-loss tuning in event vision","Context-aware thresholds make event vision sparse without tuning","Dynamic thresholds replace sparsity losses in event-vision nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3677,"prompt_tokens":710,"completion_tokens":2967,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":454,"completion_tokens_details":{"reasoning_tokens":2887}},"tokens_in":454,"tokens_out":2967,"duration_ms":25131,"temperature":1.0,"reasoning_tokens":2887,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T15:27:32.898914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a trained CSSL-SEED on a low-contrast, small-object subset of an event-detection benchmark and compare it with the same architecture with all learned thresholds forced to zero at inference. If the dense version's mAP is materially higher, the hard threshold is discarding task-critical features; if it is not, the masking is information-preserving on that distribution.","supporting_citations":[{"cited_title":"Event-based optical flow on neuromorphic processor: Ann vs. snn comparison based on activation sparsification,","cited_arxiv_id":null,"evidence_quote":"Provides the recurrent optical-flow architecture and the activation-sparsification baseline that CSSL-EV-FlowNet is compared against."},{"cited_title":"Efficient recurrent architectures through activity sparsity and sparse back-propagation through time,","cited_arxiv_id":null,"evidence_quote":"Supplies the surrogate-gradient scheme and event-recurrent training method used to backpropagate through the Heaviside mask."},{"cited_title":"Learning to detect objects with a 1 megapixel event camera,","cited_arxiv_id":null,"evidence_quote":"Provides the 1Mpx event-camera detection dataset and the evaluation protocol used for the detection experiments."},{"cited_title":"Recurrent vision transformers for object detection with event cameras,","cited_arxiv_id":null,"evidence_quote":"Provides the strong transformer-based event-detector baselines whose synaptic-operation counts CSSL is compared with."},{"cited_title":"Minimal gated unit for recurrent neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the Minimal Gated Unit that CSSL uses as its primary recurrent cell."},{"cited_title":"Deja vu: Contextual sparsity for efficient LLMs at inference time,","cited_arxiv_id":null,"evidence_quote":"Motivates input-dependent contextual sparsity, the general idea CSSL adapts to event-based vision."}],"review_version":1}