A Theoretically Grounded Spiking Neural Network Architecture for Real-Time Intracortical Signal Processing: Epistemological Foundations, Mathematical Guarantees, Computationally Verified Results, and Falsifiable Predictions for Closed-Loop Brain–Computer Interfaces

Read full text Summarise with VIRAT
Authors: Sami Rashid Mohammed Shibah
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Epistemological position. This work adopts a critical-rationalist stance (Popper, 1959): every theoretical claim is stated as a conjecture subject to empirical falsification, with quantitative rejection thresholds fixed a priori. Every mathematical guarantee is derived from explicitly stated axioms with fully written proofs, and every simulated quantitative result reported in this manuscript is regenerated by the self-contained, seeded Python implementation embedded in Appendix C, so that every number in every table and figure caption is independently reproducible by re-running the manuscript itself.Motivation. Closed-loop intracortical brain–computer interfaces (iBCIs) require decoding algorithms that jointly satisfy millisecond-scale temporal fidelity, robustness to non-stationary neural statistics, calibrated uncertainty for risk-aware control, and power budgets compatible with fully implantable hardware. Existing decoders typically sacrifice one property for another.Theoretical contributions. We present a five-part mathematical framework: (i) a generalized linear model (GLM) of spike generation proven absolutely continuous with respect to a Poisson reference measure (Theorem 3.1); (ii) a stochastic leaky integrate-and-fire (LIF) hidden layer with proven well-posedness and exponential moment bounds (Lemma 3.3); (iii) a weight-dependent spike-timing-dependent plasticity (STDP) rule with a corrected, well-posed depression term and a complete Lyapunov convergence proof (Theorem 3.4); (iv) a Bayesian Monte Carlo (MC)-dropout readout with a finite-sample calibration bound (Theorem 3.6); and (v) a Sobol global-sensitivity framework with a stated statistical-consistency guarantee (Proposition 3.9).Computationally verified results. On a 300-second synthetic 3-D center-out reaching task (96 GLM-generated afferents calibrated to published intracortical statistics), the proposed architecture achieves position/velocity NRMSE of 0.327/0.255, a median per-inference latency of ≈0.04 ms (versus ≈1.4 ms for the Kalman filter, a ≈35× reduction; wall-clock latency is inherently more variable across runs and hardware than the seeded accuracy/calibration/operation-count figures, and is reported here as a median over repeated inference passes — see Section 4.5), an expected calibration error of 0.083 with 97.2% empirical coverage of nominal 95% credible intervals, and an event-driven synaptic-operation count of ≈616 operations per 10 ms inference window — within 7% of the simplest (ridge-linear) baseline's operation count and 10.6–1,453× fewer operations than the remaining independently implemented baseline decoders (feedforward network, echo-state network, Kalman filter), at comparable or better decoding accuracy on most axes and substantially better calibration and latency than the closest (Kalman) baseline. We report these results without embellishment: on raw position/velocity NRMSE the proposed decoder is competitive with, not uniformly superior to, the simplest linear baseline; its principal, reproducibly demonstrated advantages are latency, operation-count-based energy efficiency, and calibrated uncertainty, and we state this trade-off explicitly rather than overstate accuracy gains.Global sensitivity and robustness. A Sobol variance-based sensitivity analysis (Saltelli sampling, Jansen total-order estimator, N = 16 base samples, 208 model evaluations, reduced-order 4-second proxy horizon) identifies the LIF firing threshold and the synaptic time constant as jointly dominant, strongly interacting drivers of decoding error (ST = 0.72 and 0.71 respectively, with a second-order interaction term SVth,τs = 1.41 that exceeds either individual total-order effect), followed by a closely clustered second tier comprising channel noise, MC-dropout rate, membrane time constant, and STDP potentiation rate (ST ≈ 0.35 for all four); only the refractory period and the readout learning rate are effectively inert (ST < 0.001) over the tested ranges. We report this as a screening-level, reduced-order analysis, state explicitly the statistical limitations of N = 16 (Proposition 3.9), and emphasize — rather than understate — that the dominant interaction term means the underlying sensitivity landscape is not well approximated by an additive, single-parameter picture.Falsifiable predictions. Three quantitatively bounded, pre-registered conjectures (P1–P3) are stated with explicit rejection criteria for future empirical replication on real intracortical recordings.Scope. This is a theoretical and computational study. No human or animal data were collected or used; all neural data are synthetic, generated from first-principles statistical models calibrated to published population statistics from the BrainGate2 and NeuroTycho consortia. Clinical validation is explicitly identified as necessary future work in the dedicated Roadmap (Section 8).

Fields of Study & Subjects

Neural dynamics and brain function EEG and Brain-Computer Interfaces Neuroscience and Neural Engineering

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
27 March 2026
Language
EN
Document type
Dataset