Self-Dropping Neuron Mechanisms

• Self-Dropping Neuron is a family of mechanisms where neurons autonomously regulate their activity via suppression, resetting, or pruning based on their state and reliability.
• Key implementations include selective spike suppression in single-timestep SNNs, reliability-aware remapping in cross-session decoding, and intrinsic self-reset in spintronic neuromorphic devices.
• These methods leverage dynamic thresholding, Bayesian optimization, and feedback control to enhance information capacity and reduce energy consumption in neural computations.

Searching arXiv for the cited papers to ground the article and confirm metadata. arxiv_search query: (Xu et al., 1 Aug 2025) Self-dropping neuron denotes a family of mechanisms in which a neuron, neural unit, or recorded neural channel is selectively suppressed, reset, remapped, or structurally removed on the basis of its own state, inferred reliability, or dynamical history. The term is used most explicitly for the Self-Dropping Neuron in the single-timestep spiking framework SDSNN, where dynamic threshold adjustment and selective spike suppression are used to increase information-carrying capacity per spike in a one-step inference regime (Xu et al., 1 Aug 2025). Closely related interpretations appear in cross-session neural decoding, where unreliable recorded neurons are effectively ignored or remapped rather than propagated unchanged to the decoder (Jude et al., 2022), in spintronic neuromorphic hardware with intrinsic self-reset via synthetic antiferromagnetic coupling (Sekh et al., 2 Feb 2026), and in dynamical neuroscience, where delayed self-feedback or autocorrelation-based control can suppress or tune a neuron’s own activity (Wang et al., 2015, Moraes et al., 2022). The expression therefore names a concept with multiple technical instantiations rather than a single canonical model.

1. Terminological scope and principal meanings

Across the cited literature, “self-dropping” refers to several distinct operations that share a common motif: a neuron or neuron-like element does not remain a passive recipient of inputs, but instead undergoes a selective self-limiting, self-resetting, or self-pruning process. This suggests that the term is best understood as an umbrella category spanning computational neuroscience, machine learning, neural decoding, and spintronic neuromorphic devices (Xu et al., 1 Aug 2025, Jude et al., 2022, Sekh et al., 2 Feb 2026, Wang et al., 2015).

Domain	Mechanism	Role of “dropping”
Single-timestep SNNs	Dynamic threshold adjustment and selective spike suppression	Redundant spikes are suppressed
Cross-session neural decoding	Self-supervised identification of consistent neuron ensembles	Unreliable neurons are ignored or remapped
Spintronic neuromorphic hardware	Intrinsic self-reset via SAF coupling	The device returns to resting state automatically
Dynamical neuroscience	Delayed self-feedback or autocorrelation-guided control	A neuron can suppress or retune its own firing
Network simplification and regularization	Group sparsity, targeted stripping, or deterministic thinning	Entire neurons or synapses are removed or masked

A recurrent source of confusion is that these mechanisms are not equivalent to standard dropout. In CAPTIVATE, the idea is explicitly not framed as a literal dropout layer but as a learned mechanism that ignores, replaces, or remaps neurons that are unreliable across recording sessions (Jude et al., 2022). In NSDropout, the resulting thinned network cannot be “unthinned” for testing (Shunk, 2022). In DropNeuron, neurons disappear because all incoming or outgoing weights are driven to zero by group regularization rather than being randomly masked (Pan et al., 2016).

2. Selective spike suppression in single-timestep spiking neural networks

The most literal use of the term appears in “SDSNN: A Single-Timestep Spiking Neural Network with Self-Dropping Neuron and Bayesian Optimization” (Xu et al., 1 Aug 2025). The central problem is that conventional Spiking Neural Networks, especially those based on the Leaky Integrate-and-Fire neuron, distribute information across many time steps. When inference is compressed into one timestep, a standard spike rule can yield too few spikes, too many redundant spikes, dead neurons, and incorrect state updates in deep networks. The Self-Dropping Neuron is introduced to make spike emission selective rather than purely threshold-crossing based.

The membrane update retains a LIF-style form,

$$u^{(t,n+1)} = \gamma u^{(t-1,n+1)} - V_{th}\cdot o^{(t-1,n+1)} + x^{(t,n)} \tag{6}$$

but the spike-generation rule is altered so that the neuron fires only when the membrane potential exceeds threshold and is lower than the previous time-step membrane potential:

$$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$

In the paper’s wording, the neuron generates spikes “when the membrane potential exceeds the threshold and continues to rise until it starts to decay.” This is the operative sense of self-dropping: spikes during the continuing rise are suppressed, and firing occurs on the descending side of the membrane trajectory. The spike magnitude is determined by $\left\lfloor u/V_{th}\right\rfloor$ and then clipped by a maximum firing limit, which increases information-carrying capacity beyond a strictly binary event (Xu et al., 1 Aug 2025).

To reduce dead neurons, the threshold is made time dependent:

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$

The paper also uses a localized surrogate gradient,

$$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$

with $a=1$, and states that gradients are computed only on the final time step in the single-timestep setting. This avoids time-domain BPTT-style gradient propagation and is part of the broader claim that memory complexity drops from $O(T)$ to $O(1)$ in the authors’ description (Xu et al., 1 Aug 2025).

The neuron mechanism is embedded in a single-timestep pipeline consisting of a convolutional layer, an SD spiking layer, max pooling, another SD spiking layer, and a voting layer. Layer-wise time-step configurations are chosen by Bayesian optimization through three stages: Global shared step search, Initial target-epoch search, and Final target-epoch search. The reported accuracies are 93.72% on Fashion-MNIST, 92.20% on CIFAR-10, and 69.45% on CIFAR-100, all using only single-timestep spikes. Relative to traditional multi-timestep SNNs employing the LIF model, the reported energy reductions are 56%, 21%, and 22%, respectively. On Fashion-MNIST and CIFAR-10, the SD neuron is reported to exceed the LIF baseline; on CIFAR-100, the SD result (69.45%) remains below the LIF result (70.42%), so the performance claim is comparable or even superior overall rather than uniformly dominant (Xu et al., 1 Aug 2025).

3. Reliability-aware neuron suppression in cross-session neural decoding

A different, but closely related, interpretation appears in “Capturing cross-session neural population variability through self-supervised identification of consistent neuron ensembles” (Jude et al., 2022). Here the problem is chronic neural recording drift: individual neurons can change firing properties over time, electrodes can shift, units can disappear or reappear, spike sorting and SNR changes can alter apparent identities, and neuron ordering in the input vector can change across sessions. Since population decoders assume consistent correspondence between input channels and the underlying latent manifold, this instability causes a decoder trained on day 0 to encounter a different coordinate system on later days.

The central assumption is that single-neuron identity is unstable while the population-level latent manifold governing behavior is relatively stable. CAPTIVATE treats session-to-session instability as a perturbation problem. A self-supervised neuron locator RNN is first trained on a single session using perturbed versions of each trial. The perturbations are designed to mimic cross-session changes:

• A) replace a neuron’s spike train with a randomly generated neuron
• B) add spikes to existing neurons
• C) remove spikes from existing neurons
• D) delete entire neurons and insert random ones to keep dimensionality fixed
• E) jitter all spike trains in time by $-30$ to $+30$ ms
• F) shift the entire neuron ordering by a constant fraction of the population

For each neuron in a perturbed trial, the locator predicts the original position in the unperturbed day-0 recording or an extra “randomly generated neuron” class. Conceptually, it learns a consistency mask: neurons that can be mapped back to the day-0 ensemble are treated as trustworthy, whereas neurons that cannot be matched are treated as unfamiliar and can be ignored or remapped. This is the paper’s strongest connection to a self-dropping-neuron interpretation (Jude et al., 2022).

After the locator is trained, its weights are frozen and its RNN activations condition a second model, an LFADS-inspired sequential variational autoencoder. This second stage reconstructs original unperturbed neural activity from perturbed input and produces a latent state for later behavior decoding. The objective combines Poisson reconstruction, KL regularization toward $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$0, and an explicit latent alignment term,

$$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$1

which forces different perturbations of the same trial to agree in latent space. The paper explicitly states that $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$2 and $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$3 form the usual VAE ELBO, while $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$4 accelerates training by making perturbed views of the same trial agree in latent space (Jude et al., 2022).

Experiments use publicly available M1 recordings from two monkeys performing an 8-target center-out reach task, with spikes binned at 10 ms, behavior sampled at the same resolution, only successful trials used, all trials aligned to movement onset and cut to the shortest reach duration, training on day 0, and testing on later unseen days. Monkey C has 55 neurons across sessions and 5 days total in the main analysis, with decoding remaining strong up to 8 days into the future in the longer analysis. Monkey M has 17 neurons across sessions and 5 days total. With the optimal total perturbation rate of 40%, the neuron locator achieves 85% accuracy for Monkey C and 93% for Monkey M, and still reaches 76% accuracy at 80% perturbation on an artificially heavily perturbed session. For held-out day-1 session decoding, full CAPTIVATE achieves mean $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$5 of 0.84 ± 0.02 for Monkey C and 0.86 ± 0.03 for Monkey M, outperforming CAPTIVATE-noLoc, vanilla LFADS, and a simple GRU baseline. The ablation study reports that removing perturbations mimicking neuron replacement, deletion, and reordering harms performance the most, which the paper interprets as evidence that these are dominant sources of session variability (Jude et al., 2022).

4. Intrinsic self-reset in spintronic artificial neurons

In “Spin splitting torque enabled artificial neuron with self-reset via synthetic antiferromagnetic coupling,” self-dropping refers to intrinsic reset in a physical neuron device rather than to spike suppression or pruning (Sekh et al., 2 Feb 2026). The reported neuron is an altermagnet/Synthetic Antiferromagnetic Coupling based spintronic neuron that uses out of plane spin polarized spin-splitting torque to eliminate the need for an external in-plane magnetic field $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$6. The device is built on epitaxial RuO$$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$7(101) with an added Pt layer and a synthetic antiferromagnetic stack of soft and hard ferromagnetic layers coupled through Ru spacer layers.

The key physical idea is that the neuron integrates incoming electrical stimulation by current-induced switching of a magnetic soft layer and then returns automatically to its initial state when stimulation stops. The restoring force is the built-in interlayer exchange field $$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$8 produced by synthetic antiferromagnetic coupling. In this sense, “self-dropping,” “self-resetting,” and “intrinsic reset” all describe the same phenomenon: the neuron fires under current and drops back to its resting state when the current is removed, without a separate reset pulse, an auxiliary circuit, or an external in-plane field (Sekh et al., 2 Feb 2026).

Field-free operation is enabled by the spin-splitting effect in altermagnetic RuO$$o^{(t,n+1)} = \begin{cases} \left\lfloor \frac{u^{(t,n+1)}}{V_{th}} \right\rfloor, & V_{th}< u^{(t,n+1)} < u^{(t-1,n+1)} \ 0, & \text{otherwise.} \end{cases} \tag{7}$$9(101). The paper states that the Néel vector is canted toward the $\left\lfloor u/V_{th}\right\rfloor$0-direction, so an in-plane charge current produces an out-of-plane spin polarization $\left\lfloor u/V_{th}\right\rfloor$1, which in turn exerts a damping-like spin torque of the form

$\left\lfloor u/V_{th}\right\rfloor$2

The device also includes Pt, which generates an $\left\lfloor u/V_{th}\right\rfloor$3 spin polarization via the conventional spin Hall effect. The design therefore uses two cooperative spin sources: Pt supplies a conventional damping-like spin-orbit torque, while RuO$\left\lfloor u/V_{th}\right\rfloor$4 provides the field-free symmetry breaking needed for deterministic switching at

$\left\lfloor u/V_{th}\right\rfloor$5

The same SAF coupling that opposes switching also supplies the reset force once the current is removed (Sekh et al., 2 Feb 2026).

Experimentally, the isolated soft ferromagnetic layer shows monotonic increase of anomalous Hall voltage with current and 100% field-free magnetization switching at the highest current. In the full SAF stack, switching is harder, so a small out-of-plane assistance field $\left\lfloor u/V_{th}\right\rfloor$6 is applied during measurements, but it is kept smaller than the intrinsic switching field of the isolated soft layer. For the best-performing sample, the paper reports $\left\lfloor u/V_{th}\right\rfloor$7 Oe and minimum $\left\lfloor u/V_{th}\right\rfloor$8 Oe. The device was cycled over 6 repetitions, with nearly identical integration/reset timing across cycles; near threshold there are small stochastic variations due to thermal fluctuations. Micromagnetic simulations of a domain-wall device show stepwise domain-wall motion under current and return to the original position when current is turned off. The paper studies current densities from

$\left\lfloor u/V_{th}\right\rfloor$9

reports a representative interlayer exchange energy density

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$0

and examines pulse durations with

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$1

For spiking-neural-network modeling, the device is mapped to an adaptive leaky integrate-and-fire equation,

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$2

and the reported SNN accuracies are 95.99% on MNIST and 94.36% on N-MNIST (Sekh et al., 2 Feb 2026).

5. Delayed self-feedback, self-suppression, and self-tuned criticality

In theoretical neuroscience, self-dropping is closely associated with self-feedback loops that modify a neuron’s own future state. “Firing dynamics of an autaptic neuron” reviews autapses, synapses that connect a neuron to itself, and treats them as delayed self-feedback introducing an additional time scale to neuronal dynamics (Wang et al., 2015). For electrical autapses, the current is

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$3

and for chemical autapses,

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$4

In a conductance-based form,

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$5

with

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$6

These delayed terms let the neuron act on itself through its own past activity.

The review reports that autaptic feedback can drive transitions among quiescent/resting state, periodic firing, chaotic firing, and chaotic bursting; inhibitory autapses can push a Hindmarsh-Rose neuron into a resting state; and an HH neuron can undergo spiking death, meaning that delayed self-input drives an initially firing neuron back to a fixed point. The paper also reports modifications of $V_{th} = \frac{V_{th0}}{t}. \tag{8}$7 mode-locking under sinusoidal drive, periodic modulation of firing frequency by delay, and structured changes in interspike interval distributions. Electrical and excitatory chemical autapses can either advance or delay the next spike, whereas inhibitory chemical autapses only delay it. The filtering interpretation is likewise explicit: without an autapse the neuron tends to act as a low-pass filter; with electrical or excitatory autapse and sufficiently long delay, it can become a band-pass filter; inhibitory autapse remains essentially low-pass. In this literature, self-dropping is best understood as self-regulation and, in some regimes, self-suppression (Wang et al., 2015).

A related but distinct framework appears in “Self Tuned Criticality: Controlling a neuron near its bifurcation point via temporal correlations” (Moraes et al., 2022). Here a single neuron is controlled near a bifurcation point by a feedback rule that measures temporal autocorrelation and modifies a control parameter so as to maximize that correlation. For the Chialvo map, the control law is

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$8

$V_{th} = \frac{V_{th0}}{t}. \tag{8}$9

$$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$0

and for the FitzHugh–Nagumo model the same structure is used with $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$1 as the control parameter and $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$2 as the monitored coefficient:

$$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$3

The paper notes $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$4 and uses very small Gaussian noise, with variance $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$5 for the map and $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$6 for the FHN model, so that slowing down near the bifurcation can be detected through increased autocorrelation.

For the Chialvo map,

$$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$7

with $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$8, the control is run for $$h(u) = \frac{1}{a}\,\text{sign}\left(\left|u - V_{th}\right| < \frac{a}{2}\right), \tag{10}$$9 adaptation steps, each based on $a=1$0 map iterations after discarding transients. Because the system has hysteresis and bistability, the control can converge to two nearby asymptotic values of $a=1$1. For the FitzHugh–Nagumo model,

$a=1$2

$a=1$3

with $a=1$4, $a=1$5, $a=1$6, $a=1$7, and $a=1$8, the feedback drives the system to a narrow region near the Hopf bifurcation from a stable fixed point to a limit cycle. This suggests a different sense of self-dropping: the neuron is made to “drop” itself toward the edge of oscillation by following its own temporal correlations (Moraes et al., 2022).

6. Relation to neuron dropping, synaptic stripping, and deterministic thinning

Several machine-learning methods are closely related to the self-dropping-neuron idea even when the term itself is not primary. “DropNeuron: Simplifying the Structure of Deep Neural Networks” proposes regularizers that support a mechanism of dropping neurons during training by encouraging all incoming or all outgoing weights of a neuron to become zero (Pan et al., 2016). For a neuron $a=1$9, the relevant groups are the incoming column $O(T)$0 and outgoing row $O(T)$1; a neuron can be dropped if either is exactly zero. The two principal penalties are

$O(T)$2

$O(T)$3

and the objective is the task loss plus these terms, with optional $O(T)$4 regularization. After training, weights with absolute value below a threshold such as $O(T)$5 are removed. In sparse linear regression with input dimension $O(T)$6, sparsity $O(T)$7, one hidden layer with 5 neurons, and linear activation, dropout plus pruning yielded test NMSE around 0.54, whereas DropNeuron yielded test NMSE around 0.00036, reduced total weights to about 54.11% of the original weight count, reduced neurons from 26 to 4 active units total, and achieved compression rate about 35×. On LeNet-5 for MNIST classification, $O(T)$8 + DN + pruning reached 99.07% accuracy with compression rate 67.04 (Pan et al., 2016).

“Synaptic Stripping: How Pruning Can Bring Dead Neurons Back To Life” addresses dead ReLU neurons by removing problematic incoming connections during training (Whitaker et al., 2023). A neuron is treated as dead when the sum of its activations over the validation set is zero. At the end of each epoch, the algorithm evaluates the network on a validation set, identifies dead neurons, and prunes 10% of the most negative weights feeding into each dead neuron; if a neuron has already been stripped, it prunes 10% of the remaining weights. The logic is that dead neurons contribute nothing functionally, so targeted pruning of their most negative incoming weights can shift the pre-activation upward and allow the unit to become active again. The paper reports that on small-scale Vision Transformer experiments, Synaptic Stripping increases capacity by about 15% to 33% while reducing total parameter counts by 10% to 15%. The connection to self-dropping is conceptual rather than literal: the neuron does not autonomously decide to drop its own synapses; the training algorithm does so after detecting death (Whitaker et al., 2023).

“Neuron-Specific Dropout: A Deterministic Regularization Technique to Prevent Neural Networks from Overfitting & Reduce Dependence on Large Training Samples” proposes a deterministic, class-aware thinning mechanism (Shunk, 2022). NSDropout compares the average values produced by each neuron for each class on training and validation data, identifies neurons farthest from the validation-class average, drops a fixed proportion $O(T)$9, and then propagates the masked output

$O(1)$0

Unlike standard dropout, the thinned networks cannot be “unthinned” for testing; the paper states that the best test results were obtained when the last used mask was applied. The reported results are 0% error on MNIST, 0.19% error on Fashion-MNIST, and 7.72% error on CIFAR-10. The paper also reports that with 500 images total, NSDropout reached 99.92% accuracy, with 750 images it reached its first perfect accuracy, and with only 50 images total it still exceeded 90% accuracy. Across the paper, the mechanism is explicitly differentiated from random dropout because it removes units deemed harmful by training-versus-validation comparison rather than by stochastic masking (Shunk, 2022).

Taken together, these methods show that “self-dropping neuron” can mean at least four non-equivalent operations: selective spike suppression, reliability-aware channel rejection, intrinsic physical reset, and structural neuron or synapse elimination. The common principle is that activity is not simply propagated whenever it is present; instead, a neuron or neuron-like element is conditionally suppressed, rerouted, reset, or removed so that the larger system retains useful dynamics, stable decoding, or efficient computation.