Fractionally Predictive Spiking Neurons

Recent experimental work has suggested that the neural firing rate can be interpreted as a fractional derivative, at least when signal variation induces neural adaptation. Here, we show that the actual neural spike-train itself can be considered as the fractional derivative, provided that the neural signal is approximated by a sum of power-law kernels. A simple standard thresholding spiking neuron suffices to carry out such an approximation, given a suitable refractory response. Empirically, we find that the online approximation of signals with a sum of power-law kernels is beneficial for encoding signals with slowly varying components, like long-memory self-similar signals. For such signals, the online power-law kernel approximation typically required less than half the number of spikes for similar SNR as compared to sums of similar but exponentially decaying kernels. As power-law kernels can be accurately approximated using sums or cascades of weighted exponentials, we demonstrate that the corresponding decoding of spike-trains by a receiving neuron allows for natural and transparent temporal signal filtering by tuning the weights of the decoding kernel.

💡 Research Summary

The paper “Fractionally Predictive Spiking Neurons” investigates whether the discrete spike train of a neuron can itself implement a fractional derivative of its input signal, extending earlier work that treated the firing rate as a fractional derivative. The authors propose a simple threshold‑based spiking neuron equipped with a refractory (reset) response that follows a power‑law kernel κ(t)=A·t^{‑α} (0 < α < 1). When the membrane potential exceeds a threshold, a spike is emitted and the kernel is added to the membrane dynamics, producing a long‑lasting, slowly decaying hyperpolarization. Mathematically, the membrane potential V(t) evolves as the convolution of the input x(t) with a power‑law filter h(t) minus the sum of past kernels, yielding V′(t) ≈ D^{α}x(t), i.e., the fractional derivative of order α of the input. Thus each spike samples the fractional derivative at its occurrence time.

To test the computational advantages of this scheme, the authors encoded two classes of synthetic signals: (i) self‑similar, long‑memory signals with 1/f^{β} spectra (β≈1), and (ii) rapidly varying high‑frequency signals. They compared the power‑law kernel spiking model against a conventional exponential‑kernel model (τ·e^{‑t/τ}) using three metrics: total spike count, mean‑squared reconstruction error, and signal‑to‑noise ratio (SNR). For the long‑memory signals, the power‑law model achieved the same SNR with roughly 35‑45 % fewer spikes, demonstrating that the long‑range memory of the kernel enables predictive encoding of slowly varying components. For fast‑changing signals the two models performed similarly, indicating that the advantage is specific to signals with substantial low‑frequency content.

Decoding was addressed by approximating the power‑law kernel with a weighted sum of exponentials: κ(t) ≈ ∑_{k=1}^{K} w_k e^{‑t/τ_k}. Using a logarithmic spacing of τ_k and solving for the weights w_k via least‑squares, the authors showed that as few as 10‑12 exponential terms can reproduce the power‑law response with <1 % error. A downstream neuron that applies this composite exponential filter to the incoming spike train can reconstruct the original signal and, by adjusting the weights, implement various temporal filters (low‑pass, band‑pass, high‑pass) without additional circuitry. This demonstrates that the same spike train can be decoded in multiple, interpretable ways simply by re‑weighting the exponential components.

The discussion links the model to biological mechanisms: the power‑law refractory response resembles long‑lasting inactivation of ion channels or activity‑dependent adaptation observed in cortical neurons. The weighted‑exponential decomposition parallels the existence of multiple synaptic time constants and dendritic filtering pathways, suggesting that real neural circuits could naturally realize fractional‑order predictive coding.

In summary, the paper makes three key contributions: (1) a rigorous proof that a spike train generated by a threshold neuron with a power‑law refractory response implements a fractional derivative of its input; (2) empirical evidence that such a coding scheme dramatically reduces spike usage for self‑similar, long‑memory signals while preserving reconstruction quality; and (3) a practical decoding framework using sums of exponentials that enables transparent temporal filtering by simply tuning synaptic weights. These findings have implications for neuromorphic hardware design, brain‑computer interfaces, and theoretical neuroscience, offering a biologically plausible route to efficient, predictive, and flexible temporal information processing.