Neural Noise- In leaky integrate-and-fire (LIF) spiking networks, performance and stability depend on neuron hyperparameters and inference-time state.1
- An operational manifold is a region in neuron hyperparameter space where spiking activity remains stable and usable; it clarifies accuracy–energy trade-offs and robustness for deployable SNNs.1 1
Weekly enrichment (2026-07-20)
- The study (Mazurek, Caputa, Maj & Wielgosz; AGH University of Krakow, Sano Centre, and Brookhaven National Laboratory), published as an original research article in Frontiers in Neuroscience on 18 February 2026, formalizes the “operational manifold”: a contiguous band in leaky integrate-and-fire (LIF) neuron hyperparameter space where spiking activity stays between silent and saturated regimes while task accuracy is preserved.2 3
- The manifold is mapped mainly over two parameters, the membrane time constant (τm) and firing threshold (Vth), via uniform grid sweeps of roughly 10×10 points spanning τm ∈ [1.001, 5] ms and Vth ∈ [0.01, 3] V (extended to 10 ms/10 V for the MLP-SNN on MNIST because activity did not collapse within the narrower range).2
- Architectures evaluated include an MLP-SNN (512-256-64), a VGG-like ConvSNN, SpikingResNet18, and SpikingVGG11 plus recurrent variants, trained with arctangent surrogate gradients and Adam over T=10 timesteps on MNIST, CIFAR-10, the UCF11 video dataset (11 classes), and an event-stream N-MNIST/MNIST-DVS benchmark (“EventMNIST”).2
- Two failure modes bound the manifold: a silent regime at low τm and high Vth (vanishing activity, lost information propagation) and a saturated regime at high τm and low Vth (near-tonic, epilepsy-like firing that inflates energy); inside the diagonal balanced band many settings reach near-maximal accuracy while firing far fewer spikes.2
- The authors define two composite efficiency scores, BESλ (λ weights energy versus accuracy) and EASβ (an Fβ-like score where β>1 stresses accuracy), to expose accuracy–energy frontiers; figures used λ=0.5 and β=5, with EAS producing sharper optima and BES highlighting broader stable regions, and energy approximated by a synaptic-operation (SOP) proxy.2
- Comparing inference-time state policies, hard reset gave higher accuracy on static i.i.d. inputs and reached near-saturated accuracy by roughly T≈3–4, whereas carrying membrane state introduced cross-sample interference that only faded as the integration window grew, favoring carry-style policies for streaming neuromorphic sensing.2
- For robustness, inputs were progressively frame-permuted until predictions flipped (collecting up to 2000 clean/perturbed activity pairs); leaving the manifold raised inter-neuronal spike-train correlations and synchrony, most strongly in deeper layers.2
- Higher-order statistics of the spike-train correlation distribution (skewness, kurtosis, tail mass, and counts of coefficients above 0.75 and 0.9) acted as label-free drift/health indicators, and an XGBoost classifier plus mutual-information analysis confirmed these features best separated clean from noisy inputs.2
- A hardware takeaway is that within the balanced band SNNs become “compute-light but memory-bound,” so on event-driven neuromorphic ASICs the dominant cost shifts from synaptic operations to SRAM residency of neuron state and spike-queue management.2
- Reported limitations: the manifold’s exact location and width are dataset- and training-recipe-dependent (e.g., surrogate-gradient choice), energy figures are proxy estimates rather than silicon measurements, and the SpikingResNet18/UCF11 case peaked at only about 40% accuracy since training was not tuned for that task.2