Spike Reserving: A Cross-Domain Design Pattern

• Spike Reserving is a design pattern that selectively preserves vital signal attributes—such as spike morphology, timing, or extrema—across various transformations and compressions.
• It is applied in unsupervised spike sorting, accelerated SNN inference, adversarial retiming, continual learning, and low-bit communication to maintain critical information integrity while reducing computational costs.
• The method strategically identifies and reserves key features before costly operations, ensuring robust performance even under compression, energy constraints, or adversarial settings.

Searching arXiv for the cited works and related usage of “Spike Reserving” to ground the article. Found the relevant arXiv papers on “Spike Reserving” and related formulations across spike sorting, SNN acceleration, adversarial retiming, continual learning, packed-neuron design, and communication quantization: (Zamani et al., 2022, Shen et al., 24 Jan 2025, Yu et al., 3 Feb 2026, Meem et al., 12 Feb 2026, Qin, 14 Mar 2026, Li et al., 4 Aug 2025), and the contrastive “spike” terminology in LLM training (Takase et al., 2023). Spike Reserving is a non-standard but increasingly reusable technical idea that appears across several research areas as a preservation operation applied before a costly or destructive transformation. In the literature considered here, it denotes preserving essential spike morphology during waveform compression for unsupervised spike sorting, preserving temporal spike information during accelerated spiking neural network inference, preserving the spike multiset under timing-only adversarial retiming, reserving activity under energy-aware spike budgets in continual learning, preserving spike-borne information through packed-neuron dynamics, and reserving numerical outliers during low-bit communication quantization (Zamani et al., 2022, Qin, 14 Mar 2026, Yu et al., 3 Feb 2026, Meem et al., 12 Feb 2026, Shen et al., 24 Jan 2025, Li et al., 4 Aug 2025). The shared structure is selective invariance: some aspect of the original signal is held fixed while another degree of freedom is compressed, shifted, aggregated, or quantized.

1. Terminological scope and common structure

“Spike Reserving” does not denote a single canonical algorithm. Instead, the phrase is used in multiple, domain-specific senses. In extracellular spike sorting, it refers to preserving essential spike shape while reducing the number of samples passed to feature extraction. In GPU-centric SNN acceleration, it denotes preserving temporal resolution by reusing a shared convolutional response across multiple leaky integrate-and-fire updates. In event-driven adversarial analysis, it denotes preserving spike counts and amplitudes while retiming timestamps. In continual SNN learning, it denotes reserving spike activity under a dataset-specific budget. In communication systems, it denotes reserving extreme values as floating-point numbers so that the remaining elements can be quantized more aggressively (Zamani et al., 2022, Qin, 14 Mar 2026, Yu et al., 3 Feb 2026, Meem et al., 12 Feb 2026, Li et al., 4 Aug 2025).

Across these uses, the object being reserved differs, but the operational pattern is stable. A costly downstream operator is identified—feature extraction, convolution, temporal dynamics, continual training under power constraints, or cross-GPU communication—and a subset of information deemed disproportionately important is exempted from the main compression or aggregation path. This suggests that “Spike Reserving” is best understood as a design pattern rather than a fixed formalism.

Domain	Reserved object	Main transformation constrained by reservation
Unsupervised spike sorting	Essential spike morphology	Non-uniform waveform compression
SNN inference acceleration	Temporal resolution or shared response over time	Temporal aggregation of convolution
Event-driven adversarial setting	Spike counts and amplitudes	Timing-only retiming
Continual SNN learning	Spike-rate headroom under energy constraints	Online training across tasks
Packed-neuron SNN design	Spike-borne temporal evidence	Sequence packing and decoding
Distributed communication	Per-group extrema	Low-bit quantization and packing

A common misconception is that the phrase always means preserving biological spikes as time events. That is not supported by the literature. In some papers the reserved entity is temporal structure; in others it is morphology, energy budget, or numerical outliers. Another misconception is that reservation implies exact reconstruction. In fact, some formulations are explicitly exact, such as capacity-preserving retiming or restoring floating-point extrema, whereas others are approximate but designed so that task-level behavior remains nearly unchanged.

2. Morphology reservation in unsupervised spike sorting

In "Efficient Approximation of Action Potentials with High-Order Shape Preservation in Unsupervised Spike Sorting" (Zamani et al., 2022), the problem is the hardware cost of feature extraction in the conventional spike sorting chain comprising detection and alignment, feature extraction and dimensionality reduction, and clustering. The paper tackles “Spike Reserving” in the sense of preserving essential spike morphology while compressing the waveform prior to unsupervised spike sorting. The approximation unit is inserted between alignment and feature extraction, and its goal is to shrink the sample count $N$ without a reconstruction step.

The method models aligned spikes with a Taylor polynomial centered at the aligned spike peak,

$$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$

and uses cascaded discrete derivatives,

$$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$

to identify high-information loci. The sampling unit enforces three criteria: preserve curvature using peaks in $SD$, connect sampled curvatures using $FD$, and capture higher-order convex components using $TD$. A robust selection uses three samples around $SD$ peaks, one sample at $FD$ peaks, and one at $TD$ peaks, retaining 22 of 66 samples and yielding 3X compression while preserving shape (Zamani et al., 2022).

The significance of reservation here is morphological rather than temporal. “High-order shape preservation” means retaining amplitude, slope, curvature, and higher-order convexity changes that define spike morphology, including peak–trough amplitude, width, rapid rise and fall slopes, and the after-hyperpolarization tail. The paper states that classification error remains almost unchanged before and after approximation across nine interleaved channels with three distinct spike morphologies per channel, and is almost zero for $\sigma_N \in \{0.05, 0.1, 0.15, 0.2\}$. For DWT feature extraction, CER is 0.01 both before and after spike approximation. Because the dominant hardware cost scales with the number of input samples, the compression reduces arithmetic and memory bandwidth; the paper reports up to 8.7X implementation-cost reduction for multiplier-heavy extractors such as uPCA, from 1,193,940 to 136,620 under the complexity metric $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$0 (Zamani et al., 2022).

The main trade-off is that the reservation mechanism depends on reliable peak alignment and derivative peak detection. The paper notes qualitative failure modes under extreme noise or misalignment, and it states that further compression beyond 3X may harm curvature preservation and clustering separability. This use of Spike Reserving is therefore a selective waveform-retention scheme optimized for downstream separability rather than signal reconstruction.

3. Temporal reservation in accelerated and adversarial SNNs

In "Collapse or Preserve: Data-Dependent Temporal Aggregation for Spiking Neural Network Acceleration" (Qin, 14 Mar 2026), Spike Reserving appears as a temporal decision problem. The paper argues that fine-grained, unstructured spike sparsity does not yield meaningful GPU speedup on SIMD hardware, and instead proposes Temporal Aggregated Convolution (TAC), which aggregates $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$1 spike frames before one convolution call. For grouped spikes $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$2, the aggregate is

$$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$3

followed by one convolution $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$4. TAC collapses time by reducing the output length from $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$5 to $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$6, which is beneficial on rate-coded data where temporal samples are redundant. TAC-TP, by contrast, preserves temporal resolution by sharing the same $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$7 across $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$8 independent LIF steps, so downstream spikes are generated at full temporal length $$f(x) \approx f(a) + f'(a)(x-a) + \frac{1}{2!}f''(a)(x-a)^2 + \frac{1}{3!}f'''(a)(x-a)^3 + \cdots ,$$9 even though convolution calls still drop by a factor of $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$0 (Qin, 14 Mar 2026).

The empirical distinction is explicitly data-dependent. On MNIST and Fashion-MNIST, TAC improves both speed and accuracy, with 13.8times speedup and +1.6% accuracy on MNIST and +5.4% on Fashion-MNIST in the abstract. On event-based DVS128-Gesture, standard TAC harms performance because temporal collapse destroys motion information, whereas TAC-TP preserves temporal resolution and achieves 95.1% accuracy versus a 96.3% baseline with 50% fewer convolution calls; standard TAC at the same convolution budget drops to 91.3%. The paper formalizes temporal preservation as keeping the output temporal dimension at every layer equal to $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$1, whereas repeated TAC yields $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$2 (Qin, 14 Mar 2026).

A different but related sense appears in "Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks" (Yu et al., 3 Feb 2026). There, reservation is adversarial and exact rather than accelerative. The input is an event grid $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$3, and the attacker retimes existing spikes under a capacity-1 threat model with budgets $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$4, $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$5, and $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$6. Feasible assignments must satisfy timeline consistency and non-overlap, and the strict placement operator preserves amplitudes and per-line counts. For each line $$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$7,

$$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$8

The paper explicitly connects this to “Spike Reserving”: each original spike is reserved and moved injectively to a new time bin on the same line, or kept in place. The adversary is therefore rate-preserving and, in the stronger per-line sense used in the details, spike-preserving (Yu et al., 3 Feb 2026).

The contrast between these two papers is instructive. In TAC-TP, reservation protects useful temporal information from being collapsed by optimization for throughput. In spike-retiming attacks, reservation protects counts and amplitudes while exploiting the vulnerability of temporal computation to timing perturbations. Both uses hinge on the claim that time itself is information-bearing in event-driven SNNs, but one exploits that fact for efficiency-aware design and the other for robustness analysis.

4. Activity and information reservation inside SNNs

In "Energy-Aware Spike Budgeting for Continual Learning in Spiking Neural Networks for Neuromorphic Vision" (Meem et al., 12 Feb 2026), Spike Reserving is cast as controlled activity allocation during continual learning. The monitored quantity is the batch-average spike rate

$$FD[n] = s[n]-s[n-1], \quad SD[n] = FD[n]-FD[n-1], \quad TD[n] = SD[n]-SD[n-1],$$9

and the framework imposes a dataset-specific target $SD$0 through the quadratic penalty

$SD$1

with a proportional controller

$SD$2

The paper interprets this as an energy-aware spike budgeting framework that integrates experience replay, learnable LIF parameters, and an adaptive spike scheduler. Reservation here means preserving spike headroom under sequential task learning so that early over-activity does not consume capacity needed by later tasks (Meem et al., 12 Feb 2026).

The modality dependence is central. On frame-based datasets, spike budgeting acts as a sparsity-inducing regularizer and reduces spike rates while improving accuracy. The paper reports MNIST spike rate reduction from 15.31% to 8.07% with accuracy increasing from 93.44±1.40% to 95.75±0.16%, and CIFAR-10 reduction from 37.50% to 30.90% with accuracy increasing from 59.50±1.58% to 61.26±0.79%. On event-based datasets, the controller instead relaxes the budget when the network is underactive; the paper reports DVS-Gesture accuracy improving from 74.48% to 91.93% while spike rate increases only from 0.48% to 0.72%. This is not preservation of a fixed spike count; it is preservation of an energy-constrained operating regime that remains compatible with plasticity across tasks (Meem et al., 12 Feb 2026).

A more architectural sense of information reservation appears in "SpikePack: Enhanced Information Flow in Spiking Neural Networks with High Hardware Compatibility" (Shen et al., 24 Jan 2025). Standard LIF uses the recursive update

$SD$3

which the paper associates with information loss due to temporal dependency and repeated reset. SpikePack instead first compresses the whole presynaptic spike matrix $SD$4 into a global membrane potential

$SD$5

with $SD$6, and then decodes the output sequence from that packed state. The paper states that this reserves the information carried by all presynaptic spikes across $SD$7 steps before any output spikes are decided. Because the entire computation can remain in the zipped domain, it achieves constant $SD$8 time and space complexity with respect to $SD$9 and is compatible with GeMM-friendly GPU execution (Shen et al., 24 Jan 2025).

These two papers share the idea that reservation can occur inside the model rather than at the input boundary. In continual learning, the reserved resource is activity budget. In SpikePack, it is temporal evidence that would otherwise be dissipated by repeated threshold-and-reset decisions. A plausible implication is that “Spike Reserving” in SNN research increasingly names strategies that convert temporally fragile computation into explicitly managed internal state.

5. Outlier reservation in low-bit communication

In "FlashCommunication V2: Bit Splitting and Spike Reserving for Any Bit Communication" (Li et al., 4 Aug 2025), the term is used outside neuromorphic computation. Here Spike Reserving is a per-group outlier-handling quantization scheme for cross-GPU communication at arbitrary bit widths. Tensor elements are partitioned into groups $FD$0 of size $FD$1, and exactly two spikes per group are reserved: the minimum and maximum,

$FD$2

with indices $FD$3 and $FD$4. These reserved extrema are transmitted as floating-point values, while the remaining elements are quantized by asymmetric uniform round-to-nearest over the shrunk range $FD$5,

$FD$6

The paper uses $FD$7 for $FD$8 and $FD$9 for $TD$0, stores indices as INT8, and stores scale and zero as BF16 by default (Li et al., 4 Aug 2025).

Reservation serves a precise quantization purpose: by removing the extrema from the quantized set, the dynamic range shrinks and the quantization step size decreases for the remaining elements. The paper emphasizes that this is especially important at 2-bit. On Llama-3-8B, RTN INT2 degrades C4 perplexity to 40.59, whereas Spike Reserving reduces it to 14.39. It also reports LMEval improvements such as Llama-3-8B INT2 RTN average 47.30 versus INT2_SR 61.80, and maximum communication speedups of 3.2$TD$1 in AllReduce and 2$TD$2 in All2All in the abstract (Li et al., 4 Aug 2025).

This meaning of Spike Reserving differs sharply from the SNN usages. The reserved “spikes” are numerical outliers rather than temporal events, and the task is communication-efficient quantization rather than neural inference. Even so, the abstraction is consistent: the few elements that most distort a global operation are exempted from the main compression path and transmitted exactly. The paper also makes the method orthogonal to bit splitting, which decomposes arbitrary bit widths into hardware-friendly units from the set $TD$3.

6. Conceptual synthesis, misconceptions, and adjacent terminology

Across these literatures, Spike Reserving is best characterized by what is held invariant. In spike sorting, essential morphology is retained while sample count is reduced. In TAC-TP, temporal resolution is retained while convolution calls are reduced. In rate-preserving retiming, counts and amplitudes are retained while timestamps are altered. In spike budgeting, an activity envelope is retained while the network adapts across tasks. In SpikePack, temporal evidence is retained before spiking decisions are emitted. In FlashCommunication V2, extrema are retained exactly while the remainder is quantized (Zamani et al., 2022, Qin, 14 Mar 2026, Yu et al., 3 Feb 2026, Meem et al., 12 Feb 2026, Shen et al., 24 Jan 2025, Li et al., 4 Aug 2025).

Several distinctions are important. First, reservation is not synonymous with sparsity. TAC was proposed precisely because fine-grained spike sparsity on GPUs did not outperform dense convolution; the advantage came from aggregation and selective temporal preservation rather than zero-skipping (Qin, 14 Mar 2026). Second, reservation is not synonymous with robustness. In adversarial retiming, exact reservation of the spike multiset makes the attack stealthier, not safer (Yu et al., 3 Feb 2026). Third, reservation is not always exact. Morphology-preserving spike approximation for sorting is “almost loss-less” in the sense that classification error remains essentially unchanged, but the method is still an approximation rather than an identity transform (Zamani et al., 2022).

The phrase should also be distinguished from unrelated uses of “spike” in machine learning. "Spike No More: Stabilizing the Pre-training of LLMs" studies loss spikes during Transformer pre-training and attributes them to gradient-norm growth governed by Jacobian spectral norms, with stabilizing conditions summarized as small sub-layers and large shortcut (Takase et al., 2023). That work concerns optimization spikes, not Spike Reserving. Its inclusion is useful mainly as a terminological contrast: “spike” in current literature may refer to event signals, waveform extrema, numerical outliers, or abrupt optimization failures, and only some of these contexts support the reserving interpretation.

Taken together, the literature suggests a unifying interpretation of Spike Reserving as selective preservation under constrained computation. A plausible implication is that the term will continue to broaden wherever a small subset of temporal, morphological, energetic, or numeric structure disproportionately determines downstream fidelity. Whether that broadening becomes a stable cross-domain concept remains unresolved, but the present corpus already shows that the underlying strategy has become technically productive across sensing, neuromorphic learning, and high-performance communication.