acceptodds
Under review as a conference paper at ICLR 2027

Fixed-Input State Dependence of Residual Transmission in Spiking Neural Networks

Abstract

Residual connections in Spiking Neural Networks (SNNs) merge branch and shortcut signals before a post-add neuron, yet identical merged input need not yield identical transmitted spikes when the receiver retains membrane state. We isolate this dependence by fixing the realized post-add merge and replacing only the state carried from earlier transitions. For additive leaky integrate-and-fire (LIF) and parametric leaky integrate-and-fire (PLIF) dynamics, we derive an exact switching criterion. At a fixed merge, two retained states produce different spikes precisely when their state contributions straddle the firing margin. SpikeDF, a temporal residual SNN for video forgery detection, provides a binary case in which shortcut-only and branch-only origins yield the same unit-valued merged input but different retained states. Nearby cross-video state replacement switches 11.5-15.3 % of deep residual outputs while changing mean output by at most 0.74 pp; propagating only those changed spikes also alters detector scores. The same near-threshold concentration of switches appears in independently trained continuous-merge models on FaceForensics++ and post-add PLIF models on DVS128 Gesture. Across eight PLIF residual sites, pre-replacement boundary occupancy closely tracks the ordering of site-level switch rates, linking switching prevalence to firing-boundary geometry. Together, these results show that a fixed residual merge and nearly unchanged mean activity can coexist with threshold-structured changes in spike identity and downstream computation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QANgOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies state-dependent residual transmission in SNNs by holding the merged input M fixed and replacing the retained membrane state R. It derives an exact threshold-crossing condition for additive LIF and PLIF neurons and examines how state pairing affects switching prevalence. Experiments on SpikeDF show that substantial spike switching can coexist with nearly unchanged mean output, while controlled spike substitution reveals site-dependent downstream sensitivity. Continuous-merge FF++ and DVS128 Gesture experiments further show that switching concentrates near the firing boundary. Separate recursive state interventions assess how retained state affects task performance, distinguishing local switching susceptibility from downstream sensitivity and trajectory-level dependence.

Critical strengths

Strengths

  1. The experimental design is grounded, and the coverage of experimental settings is broad. Thus, I think the experimental results are reliable.

  2. The authors provide mathematical and theoretical explanations about how status exchanges of LIF neurons affect the neurons' spiking patterns.

Critical weaknesses

Weaknesses

  1. In my opinion, the problem tackled by the authors is not significant. The problem is somewhat detailed, but I cannot assure that this problem significantly degrades the learning performance of residual SNNs. The authors fail to clarify how the tackled problem causes further limitations in residual SNNs.

  2. The authors measure how often the output spike of a post-add LIF neuron changes when its retained membrane state is replaced while the residual input remains fixed. However, I am not sure that these changes in spiking patterns cause negative effects. Changes in spiking patterns of LIF neurons are normal.

  3. The authors only use LIF-based residual SNNs. I think they should extend their analysis using various neuron models. For example, residual SNNs can be implemented with IF neurons [1].

  4. The authors use video forgery detection tasks to evaluate the SNNs' performance. Although forgery detection is a reasonable temporal task, the paper does not sufficiently justify why this application is particularly relevant to the proposed analysis.


[1] W. Fang, Z. Yu, Y. Chen, T. Huang, T. Masquelier, and Y. Tian, "Deep Residual Learning in Spiking Neural Networks," in Advances in Neural Information Processing Systems (NeurIPS), vol. 34, pp. 21056–21069, 2021.

Questions for authors

Questions

  1. First of all, does the observed phenomenon really cause negative effects on SNNs?

  2. Although the authors state that spiking activities are affected by residual connections, they do not show experimental results with changing the configurations of residual connections in the paper. Why didn't you change these configurations? I think it might be important to support the authors' claim.

  3. Why did you use temporal residual cues and hard pseudo-events in SpikeDF? Are they widely used pre-processing techniques?

12 comments

We appreciate the reviewer’s careful consideration of the functional role of retained history in residual SNNs. Our fixed-input analysis connects the neuron-level switching condition to what is observed in trained networks: the switch rate depends on how membrane states are paired, while the effects of those switches cannot be inferred from mean firing activity alone. At S3.1, random-donor replacement produces 13.61% switching, compared with 49.76% under the independent-marginal reference within merged-current bins. At S3.0, 13.33% of selected outputs switch even though mean activity changes by only −0.04 percentage points, and equal-count substitutions at different residual sites lead to different detector-score responses. Full recurrent-history interventions further show how the history associated with an input affects task performance. We address the reviewer’s specific questions below.

AuthorsOpenReview ↗

Q1 : Does the phenomenon have a functional effect? (W1, W2)

W1 : Retained history and task performance

We thank the reviewer for asking about consequences for the trained detector. Retaining the history associated with the current video has a substantial functional role: when we pair a video with another video’s history, FF++ validation AUC falls from 90.78 ± 0.16% to 75.90 ± 0.68% (paired change −14.88±0.53-14.88\pm0.53 percentage points), although late activity changes only from 4.80 ± 0.14% to 4.48 ± 0.37%. The contrast shows that the detector depends on the history accumulated with the current input, even when overall late firing activity changes relatively little.

Stage-local recurrent interventions further reveal that this dependence varies with the parts of the network affected:

Retained-history interventionFF++ AUC (%)Paired change (pp)Late activity (%)
Self history90.78 ± 0.16—4.80 ± 0.14
Clear Stage 149.25 ± 5.87−41.53 ± 5.822.75 ± 0.33
Clear Stage 257.79 ± 2.84−33.00 ± 3.002.54 ± 0.52
Clear Stage 390.59 ± 0.11−0.19 ± 0.181.63 ± 0.14
Clear Fusion90.51 ± 0.08−0.27 ± 0.203.04 ± 0.11
Clear Stage 3 and Fusion45.90 ± 0.59−44.88 ± 0.500.00 ± 0.00

Clearing Stage 3 or Fusion separately changes AUC by only 0.19 or 0.27 points, whereas clearing both decreases it by 44.88 points and leaves zero late activity. This non-additive task response shows why the functional role of retained history cannot be inferred by adding the effects of separate stage interventions.

W2 : Spike identity, mean activity, and downstream sensitivity

The fixed-input experiments explain a complementary part of this behavior. At S3.1, empirical random-donor switching is 13.61%, versus 49.76% for the independent-marginal reference within merged-current bins, demonstrating the influence of observed state pairing. At S3.0, the nearest-donor experiment changes 13.33% of selected Z=1Z=1 outputs while the net mean-activity change is only −0.04 points, because switches in opposite directions largely cancel.

The effects are visible downstream as well: propagating all identified S1.0 output substitutions through the recorded downstream computation changes the predicted label of 19.53 ± 1.56% of evaluated videos. With the same budget of 100 identified substitutions at each site, the relationship between local switching and detector-score response is:

SiteNearest-donor switching (%)Mean ∣Δscore∣\lvert\Delta\mathrm{score}\rvert at K=100K=100
S1.033.46 ± 0.830.1677 ± 0.0037
S3.013.33 ± 1.690.0658 ± 0.0044
S3.215.26 ± 0.750.0424 ± 0.0049

S3.2 switches more often than S3.0, yet produces a smaller detector-score change under the same substitution budget. By linking the observed state-pairing statistics to spike substitutions and detector responses, these measurements explain why neither local switching prevalence nor aggregate firing activity is sufficient on its own to characterize downstream sensitivity. The separate recurrent-history interventions then show how changes to the input–history relationship affect task performance over the full sequence.

AuthorsOpenReview ↗

Q2 : Why compare these residual configurations? (W3)

Residual depth and transmission rules

We appreciate the suggestion to vary residual configurations, which helps clarify the scope of our experimental comparisons. We vary residual depth and the rule applied after the branch–shortcut merge, while holding the wiring pattern consistent so that a topology change is not introduced alongside the post-add rule. In the separate fixed-input analysis, the realized merge itself is fixed; the independently trained rule variants share the wiring but need not produce identical merged inputs.

The four- and six-block FF++ comparisons include Static Supra-Unit, Fixed-Stateful LIF, learned RD-LIF, and Binary-OR, with an additional SEW-ADD-style operator that preserves multivalued residual outputs:

Post-add rule4-block FF++ AUC (%)6-block FF++ AUC (%)6-block late active-state density (%)
Static Supra-Unit82.23 ± 1.6573.97 ± 1.710.089 ± 0.023
Fixed-Stateful LIF90.95 ± 0.2391.31 ± 0.654.92 ± 0.75
Learned RD-LIF91.08 ± 0.3790.76 ± 0.204.80 ± 0.14
Binary-OR92.27 ± 0.6192.52 ± 0.8985.31 ± 0.46
SEW-ADD-style92.44 ± 0.5892.77 ± 0.4675.18 ± 0.25

In the six-block setting, the alternatives thus yield markedly different activity regimes and AUCs. We also retrained the models after removing the auxiliary continuous residual-statistics pathway:

Post-add rule, no auxiliary pathwayFF++ AUC (%)Late activity (%)
Static Supra-Unit75.78 ± 1.500.07 ± 0.03
Fixed-Stateful LIF90.84 ± 0.105.17 ± 0.32
Learned RD-LIF90.47 ± 0.564.95 ± 0.15

Across three paired checkpoints, Fixed-Stateful LIF and learned RD-LIF exceed Static Supra-Unit by 15.06 ± 1.57 and 14.69 ± 1.24 AUC percentage points, respectively, with positive differences in every checkpoint. These results establish the importance of the evaluated post-add transmission choices within matched residual wiring, while the fixed-input analysis isolates the effect of membrane history at a realized merge.

W3 : IF applicability and trained neuron-model coverage

Our independent trained-network tests include learned RD-LIF in SpikeDF, standard LIF with continuous FF++ merges, and PLIF with native DVS128 Gesture events. For an additive IF update without leak, the normalized pre-threshold state can be expressed as p=c(M)+h(R)p=c(M)+h(R), so two retained states switch the output at a fixed input when

min⁡(ha,hb)<1−c(M)≤max⁡(ha,hb).\min(h_a,h_b)<1-c(M)\leq\max(h_a,h_b).

The crossing condition therefore applies to additive IF analytically.

Reviewer_y4LjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript examined the effect of changing neuronal states while keeping the merged residual input fixed. This paper first derives the conditions for spike switching in a single neuron and the switch probability for the neuron population. In the experiments, building upon the SpikeDF, it investigated how the spike switching rate and downstream scores are affected by various state replacement conditions, including the source of “donor states”, the number or fraction, the “boundary occupancy” and datasets.

Critical strengths
  1. This manuscript provides a relatively comprehensive theoretical analysis, discussing the conditions for spike switching.
  2. Experiments are sufficient, covering how various factors of state replacement influence the switching and score.
Critical weaknesses
  1. The research objectives are not clear. The paper points out that neuronal spikes are influenced by the state, but it does not explain why “state replacement” is used to investigate this issue. It also does not explain how this will influence the “residual transmission”.
  2. The purpose of the experiments is unclear. These experiments lack an internal connection, and there is insufficient discussion of the observed phenomena.
  3. The role of the residual is not prominent in this paper. Although it emphasizes residual in the abstract and introduction, it mainly discusses the relationship between the state and “merged input” in the method and experiments.
Questions for authors
  1. "State-dependent" is an important characteristic of spiking neurons. What is the motivation for replacing the state of a neuron?
  2. What role does "residual" play in this paper? It seems that replacing the residual MM with a simple input has no impact on most of the contents of this paper, with the exception of the experiments described in Section 4.5.
  3. A more precise definition of “context XX” should be given (in line 190). The legend in Figure 3(a) appears to be incorrect.
5 comments

We appreciate the reviewer’s questions about the purpose of replacing retained state and the specific role of the residual connection. In SpikeDF, the branch and shortcut form the current merged event, but the post-add neuron also carries history from earlier transitions, which can change whether that same event is transmitted. Holding the realized merge fixed allows us to measure that contribution directly, before examining how the changed spikes affect the detector. The experiments below follow this relationship from the local transmission rule to network-level behavior.

AuthorsOpenReview ↗

Q1 : Why replace state, and how do the experiments connect? (W1, W2)

W1 : Motivation for retained-state replacement

A post-add neuron normally receives the current residual merge together with a membrane state accumulated from preceding transitions. Replacing only the latter, while retaining the observed merge and neuron parameters, gives us a controlled answer to the question: which spikes would change because the neuron carried a different history? The intervention is therefore not intended to simulate a typical video trajectory, it identifies the history-dependent part of transmission at an actual residual event.

In the nearest-donor replay, S3.0 exhibits 13.33 ± 1.69% switching with a net mean-activity change of only −0.04 percentage points, while S3.1 exhibits 13.26 ± 0.89% switching with a +0.23-point activity change. A separate random-donor comparison at S3.1 yields 13.61% observed switching against 49.76% under the independent-marginal reference, showing that the pairing of actual states matters beyond the neuron-level threshold formula.

W2 : How the local and task-level experiments fit together

The experiments follow the same computation at different scales. Fixed-input replay identifies spikes changed by retained history; controlled output substitution then measures their effect on the recorded downstream computation, whereas full-history interventions recompute the recurrent network and test task performance. The equal-count comparisons distinguish how frequently a site switches from how strongly those switches affect the detector:

Residual siteNearest-donor switching (%)Mean ∣Δscore∣\lvert\Delta\mathrm{score}\rvert at K=100K=100
S1.033.46 ± 0.830.1677 ± 0.0037
S3.013.33 ± 1.690.0658 ± 0.0044
S3.215.26 ± 0.750.0424 ± 0.0049

Although S3.2 has a higher local switch rate than S3.0, its detector-score response is smaller under the same substitution budget. This site-dependent sensitivity complements the activity-cancellation result: changes in transmitted spike identity may be hidden in the mean firing rate, while their downstream effect also depends on where they occur.

The recurrent-history controls connect the local analysis to task computation, with the effect of altering state depending on the stage or combination of stages involved:

Retained-history interventionFF++ AUC (%)Paired change (pp)Late activity (%)
Self history90.78 ± 0.16—4.80 ± 0.14
Clear Stage 149.25 ± 5.87−41.53 ± 5.822.75 ± 0.33
Clear Stage 257.79 ± 2.84−33.00 ± 3.002.54 ± 0.52
Clear Stage 390.59 ± 0.11−0.19 ± 0.181.63 ± 0.14
Clear Fusion90.51 ± 0.08−0.27 ± 0.203.04 ± 0.11
Clear Stage 3 and Fusion45.90 ± 0.59−44.88 ± 0.500.00 ± 0.00

Clearing Stage 3 or Fusion alone has little effect on AUC, whereas clearing both produces a 44.88-point reduction and eliminates late activity. Together, these results explain the progression of the experiments: identify history-dependent transmission, measure its downstream sensitivity, and assess the consequences of altering retained history through recurrent execution.

AuthorsOpenReview ↗

Q2 : What role does the residual connection play? (W3)

The residual connection is the source of the current input to the post-add neuron: a shortcut spike xx and a transformed-branch spike bb combine to give Z=x+b∈{0,1,2}Z=x+b\in\{0,1,2\}. When Z=1Z=1, exactly one path is active, whether (x,b)=(1,0)(x,b)=(1,0) or (0,1)(0,1). For 1/2<η<11/2<\eta<1 and h(R)∈[1−2η,1)h(R)\in[1-2\eta,1), the Z=0Z=0 event remains silent and the Z=2Z=2 event fires, so the post-add spike is

S=1[Z=2]+1[Z=1]1[h(R)≥1−η].S=\mathbf1[Z=2]+\mathbf1[Z=1]\mathbf1[h(R)\geq1-\eta].

The residual merge determines the current event, while membrane history determines whether a unit-valued event crosses threshold. For a fixed distribution of merged events, let πj=Pr⁡(Z=j)\pi_j=\Pr(Z=j) and q=Pr⁡[h(R)≥1−η∣Z=1]q=\Pr[h(R)\geq1-\eta\mid Z=1]; the corresponding binary activity fractions satisfy ρStatic=π2≤ρStateful=π2+qπ1≤ρOR=π2+π1\rho_{\mathrm{Static}}=\pi_2\leq\rho_{\mathrm{Stateful}}=\pi_2+q\pi_1\leq\rho_{\mathrm{OR}}=\pi_2+\pi_1. This makes the role of the residual connection more specific than replacing it with an arbitrary scalar input: our experiments observe the two-path merge and can distinguish its shortcut-only and branch-only sources. In a fixed-Z=1Z=1 cross-video replay, those sources have the following observed switching rates:

Source of the Z=1Z=1 eventOutput switching (%)
Shortcut-only, (x,b)=(1,0)(x,b)=(1,0)15.61 ± 1.13
Branch-only, (x,b)=(0,1)(x,b)=(0,1)6.44 ± 1.17

The two origins lead to the same value of ZZ, but the state pairs associated with their occurrences differ. The post-add neuron does not itself read an origin label once ZZ and RR are fixed, the measured difference arises from the histories accompanying the observed residual events.

We further compare the transmission rules applied after residual merging in independently trained six-block FF++ networks with matched branch–shortcut wiring:

Post-add residual ruleFF++ AUC (%)Late active-state density (%)
Static Supra-Unit73.97 ± 1.710.089 ± 0.023
Fixed-Stateful LIF91.31 ± 0.654.92 ± 0.75
Learned RD-LIF90.76 ± 0.204.80 ± 0.14
Binary-OR92.52 ± 0.8985.31 ± 0.46
SEW-ADD-style92.77 ± 0.4675.18 ± 0.25

For the first four binary-output rules, activity is the firing fraction; SEW-ADD-style instead reports the fraction of nonzero multivalued outputs. The matched-wiring comparisons exhibit markedly different activity regimes under the evaluated transmission rules, while fixed-input replay separately measures the effect of retained state within a given rule.

Reviewer_P98zOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how retained membrane state affects residual transmission in SNNs when the merged input is held fixed. It derives an exact switching condition for additive LIF and PLIF neurons and tests state replacement in video forgery detection and gesture recognition. The experiments show that substantial spike substitutions can coexist with small changes in mean activity, and examine their threshold dependence and downstream effects.

Critical strengths
  1. Holding the realized merge fixed cleanly isolates the immediate contribution of retained state. Exact replay checks strengthen confidence that the measured switches reflect the intended intervention.
  2. The experiments cover binary and continuous merges, two neuron models, and distinct tasks. The cancellation between spike substitutions and mean activity changes is a useful caution for interpreting aggregate firing statistics.
Critical weaknesses
  1. The central switching criterion follows directly from the standard thresholded neuron update and is not specific to residual connections. The paper does not clearly establish what new principle about residual computation emerges beyond this familiar state dependence.
  2. Nearest-donor matching uses video-level summaries, which do not ensure that substituted states are compatible with the recipient’s local input and history. The reported switch rates therefore establish intervention sensitivity, but their relationship to naturally occurring state variation remains uncertain.
  3. Appendix D.3 states that later membrane states are not recursively recomputed during downstream propagation, limiting what these results establish about normal recurrent computation. Count- and direction-matched random spike controls are also missing, leaving the specificity of the downstream effects unresolved.
  4. The boundary-occupancy diagnostic is supported mainly by an eight-site correlation at a single bandwidth. Per-seed stability, bandwidth sensitivity, and predictive validation across architectures are needed to establish its broader usefulness.
Questions for authors
  1. How are donor and recipient states aligned across channels, spatial positions, and time steps, and how large are the resulting local state differences after nearest-donor matching?
  2. Do the downstream effects and site rankings persist when later states evolve normally after a single intervention, and how do they compare with count- and direction-matched random spike substitutions?
5 comments

We appreciate the reviewer’s careful examination of the fixed-input analysis and the specific questions about donor alignment, downstream propagation, and threshold localization. The neuron-level crossing condition provides the starting point, our measurements show how the states actually paired in a trained residual network determine switching prevalence and how identified spike changes affect its downstream computation. In particular, substantial changes in spike identity can nearly cancel in mean activity, while equal-count substitutions have different effects across residual sites. We address the experimental protocols and their implications below.

AuthorsOpenReview ↗

Q1 : Donor alignment and observed switching (W1, W2, W4)

W1 : Observed switching beyond the threshold condition.

The crossing condition determines whether a particular pair of retained states produces different spikes at a fixed merged input; the switching rate in a trained network also depends on which states are actually paired. We therefore compared observed random-donor switching with an independent-marginal reference within site-specific merged-current bins:

SiteObserved (%)Independent (%)Reported gap (pp)
S1.038.2448.7910.55
S2.027.8049.5921.79
S3.013.9849.5235.54
S3.113.6149.7636.16
S3.215.5446.5230.97
S3.311.8435.7923.95

The reference preserves the state marginals within each merged-current bin but removes the observed dependence between paired states. At S3.1, this changes the estimated switch rate by a reported 36.16 percentage points. Thus, although the threshold condition specifies exactly when a pair switches, its frequency cannot be inferred from the threshold alone: the joint distribution of retained states determines how often the condition is met in the trained network. At S3.0–S3.2, observed rates of 13.61–15.54% remain well below the 46.52–49.76% references, so this is not confined to one residual site.

W2 : Donor alignment and local state differences.

Thank you for asking how the donor is aligned. Both states come from the same checkpoint and residual site and are compared at the same time step, channel, and spatial coordinate, with the receiver's realized merge and neuron parameters fixed. Holding the receiver's merge unchanged ensures that input changes are not conflated with the effect of replacing history; the alignment is by tensor coordinates, not by presumed semantic correspondence between videos. The nearest donor is another video selected using standardized video-level summaries of residual activity, membrane state, merged current, and threshold distance; it is not required to match the receiver's local state. We therefore measured the actual local difference ∣Δh∣=∣hdonor−hself∣|\Delta h|=|h_{\mathrm{donor}}-h_{\mathrm{self}}| across 1,769,472 sampled events (three checkpoints, six sites, 98,304 events per checkpoint–site pair):

SiteMedianP90P95Median at switches
S1.00.13310.55680.57090.4959
S2.00.00000.54220.55140.5345
S3.00.00000.46410.55070.5459
S3.10.00000.53960.54980.5457
S3.20.00000.54120.54780.5448
S3.30.00000.49500.54220.5410

These distributions make the distinction between donor selection and local state matching explicit. At S2.0–S3.3, the overall median is zero, but the upper tails remain substantial, and the median ∣Δh∣|\Delta h| among switched outputs ranges from 0.5345 to 0.5459. Many locations therefore retain the same state across paired trajectories, while the sites where spikes change typically have appreciable local differences. The matching selects a comparable donor trajectory at the video level; the local measurements show what was actually changed, without treating the donor as the receiver's natural alternative history.

W4: Threshold-boundary localization.

We measured switching relative to d=∣1−c(M)∣d=|1-c(M)|, reporting near-threshold (d≤0.10d\leq0.10) and far-threshold (d≥0.40d\geq0.40) switch rates and the share of outputs and switches within the boundary (d<0.40d<0.40):

ModelSiteNear (%)Far (%)Boundary outputs (%)Boundary switches (%)
FF++ LIFStage 136.80 ± 1.640.107 ± 0.0136.7294.31
FF++ LIFStage 236.05 ± 0.530.186 ± 0.0517.7991.92
FF++ LIFStage 337.83 ± 0.990.105 ± 0.0206.2494.04
DVS128 PLIFB136.07 ± 0.800.0818 ± 0.02273.5888.67
DVS128 PLIFB434.62 ± 1.160.1099 ± 0.02727.3693.92
DVS128 PLIFB828.62 ± 0.660.0286 ± 0.015815.0498.85

The concentration is pronounced across both input regimes. At PLIF B8, the boundary occupies 15.04% of outputs but contains 98.85% of switches; across eight PLIF sites, occupancy and switching correlate strongly (Spearman ρs=0.976\rho_s=0.976 at ϵ=0.40\epsilon=0.40).

AuthorsOpenReview ↗

Q2 : Downstream effects and recurrent propagation (W3)

We appreciate the request to separate the immediate effect of a changed spike from its consequences as the recurrent network evolves. Our first experiment identifies switches under a fixed merge and substitutes those outputs into the recorded downstream computation; a no-substitution replay reproduces the reference scores to approximately 10−610^{-6}. Because subsequent membrane states are not recomputed in this protocol, it directly measures downstream sensitivity along the recorded trajectory, with the following results across three checkpoints:

SiteK=1K=1: mean ∣Δscore∣\lvert\Delta\mathrm{score}\rvertK=10K=10K=100K=100Changed predictions, all identified spikes (%)
S1.00.0372 ± 0.00180.0974 ± 0.00470.1677 ± 0.003719.53 ± 1.56
S2.00.0203 ± 0.00080.0587 ± 0.00170.0952 ± 0.002710.16 ± 2.82
S3.00.0076 ± 0.00180.0357 ± 0.00590.0658 ± 0.00441.04 ± 0.45
S3.10.0054 ± 0.00080.0262 ± 0.00390.0528 ± 0.00641.30 ± 1.19
S3.20.0024 ± 0.00050.0181 ± 0.00220.0424 ± 0.00491.30 ± 0.45
S3.30.0028 ± 0.00020.0130 ± 0.00150.0363 ± 0.00321.56 ± 1.56

With K=100K=100 changed outputs, S1.0 produces 4.62 times the mean score change of S3.3, although the substitution count is identical. At S1.0, propagating increasing fractions of the identified changes gives a corresponding dose response:

Fraction substitutedMean ∣Δscore∣\lvert\Delta\mathrm{score}\rvertChanged predictions (%)
1%0.241 ± 0.0074.17 ± 1.04
10%0.376 ± 0.0175.10 ± 0.59
50%0.678 ± 0.02111.09 ± 2.04
100%1.030 ± 0.08719.53 ± 1.56

We separately tested fully recurrent history interventions, in which later inputs and membrane states are recomputed throughout the trajectory. Replacing a video’s retained history with another video’s history reduces FF++ validation AUC from 90.78 ± 0.16% to 75.90 ± 0.68%, while late activity changes from 4.80 ± 0.14% to 4.48 ± 0.37%. These full-history controls establish task dependence on the retained-state trajectory. The measured dose response and site ranking nevertheless show that the identified state-induced spikes have appreciable, location-dependent downstream effects.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.