A chaotic spiking network that doubles as a coin flipper

What the work claims

This is a combined modeling, benchmarking, and hardware study from Jafar Shamsi, Navid Akbari, Sonia Sennik, Aaron Gruber, and Wilten Nicola at the University of Calgary.¹ The claim has two layers. First, the same balanced-network regime that cortical models use for computation, in which excitatory and inhibitory coupling cancels on average (mean weight zero, variance scaling as 1/N), produces chaotic dynamics whose irregular spiking is statistically usable as a pseudo-random number generator, provided the network sits in the spike-chaos regime (asynchronous, Poisson-like firing) rather than the rate-chaos regime (bursting, slow rate fluctuations). Second, this generator is cheap: quantized and binarized, it fits a small low-power FPGA, and because the weights are binary, the number of distinct instantiable generators grows as 2 to the power N squared, with each instance reconfigurable on the fly by disabling neurons.¹

The authors position this as a dual-use primitive for neuromorphic hardware: a native random source that reuses the same circuits as computation, relevant to cryptography and to randomized algorithms. They are candid that formal cryptographic security is out of scope.

How it works

The network is 256 leaky integrate-and-fire neurons with random balanced connectivity, binarized to weights of plus or minus g over square root of N, and every parameter (time constants, thresholds, coupling) quantized to powers of two so all arithmetic becomes bit shifts and adds. In simulation this preserves the signature dynamics: inter-spike intervals with a coefficient of variation near 1 (Poisson-like, 0.99 plus or minus 0.25 before quantization, 0.68 plus or minus 0.16 after), and sensitivity to microscopic perturbation: delete one spike or flip one weight sign and the subsequent spike pattern diverges completely.¹

The subtle part is the readout. Simply binary-encoding which neuron fired fails the NIST battery, because a neuron's refractory period imprints detectable structure on its spike-index bits. The authors instead accumulate spike history with a running counter modulo N and read bits from a static lookup table indexed by that counter. That dynamic lookup table converts population spiking into bit streams where about half pass all 15 NIST tests and another 40% pass 14, matching the Mersenne Twister, Blum-Blum-Shub, and linear congruential generators, and beating quadratic congruential and modular-exponentiation generators, none of whose streams passed all 15 tests.¹ Across coupling and synaptic time constants, the number of tests passed correlates strongly and negatively with the mean inter-spike-interval coefficient of variation (rho = −0.8147), cleanly separating the bursty rate-chaos regime (poor bits) from asynchronous spike-chaos (good bits).

For demanding test suites the raw output still leaks: Dieharder flags short-range serial correlation in unconditioned streams, while TestU01 SmallCrush passes. Two passes of pairwise XOR whitening (parity of adjacent bit pairs, applied recursively) eliminate the Dieharder failures; three passes leave 4 flagged tests in 11,400, consistent with chance.¹

In hardware, the design fits a Lattice iCE40UP5K using 76% of slices and 25 block RAMs (125 kbit), runs on the 6 MHz internal oscillator (41 MHz was the estimated maximum), and draws 3.24 mW for the FPGA core by the vendor tool or 5.45 mW measured at the board (1.65 mA at 3.3 V), delivering 120 kilobits per second over UART. As a sanity check, 2.2 million generated numbers approximate pi as 3.1413 by Monte Carlo integration.¹

Where a skeptic should push

Start with the headline framing. This is a pseudo-random generator: a deterministic dynamical system with a hidden internal state. The authors simulated 1.1 terabytes of bits without observing a cycle and give a birthday-bound heuristic of an expected cycle time near 2 to the power 12000 for 32-bit state precision, but a heuristic is not a proof, and discrete-time discretized chaotic systems do eventually cycle. For statistical simulation workloads that is irrelevant; for key generation it matters a great deal.¹

Second, the security section is explicitly four-part and the fourth part, cryptographic security, is left open. State secrecy against reconstruction from observed bit streams, side-channel resistance, and fault-injection analysis are all future work. The side-channel argument, that parallel tree-adder power signatures are hard to invert, is plausible and unproven. The 120 kbps bandwidth is also acknowledged as low compared to other chaotic PRNGs, though sub-milliwatt ASIC versions are projected.¹

Third, the quality gate is narrower than it looks. Roughly half of seeded streams pass all 15 NIST tests, and individual weight matrices vary in quality; the authors even construct a pathological balanced matrix that produces poor bit streams while satisfying the balance conditions, by violating weight independence. Networks below about 64 neurons can fall into stable non-chaotic attractors, and the working range was validated at 128, 256, and 512 neurons with at least 128 neurons kept active.¹ So the generator is a distribution over networks and seeds, not a guaranteed-random object; quality control is per-instance.

The most load-bearing assumption is that passing NIST SP-800-22 plus TestU01 SmallCrush suffices to call the output random. These batteries test statistical null hypotheses against a broad class of defects; they are necessary, not sufficient, and say nothing about predictability given partial state knowledge. The paper never claims otherwise, but readers building on it should not upgrade "passes NIST" to "safe for cryptography" without the missing cryptanalysis.

Your organoid's noise is a resource, and a risk

For organoid intelligence, this paper delivers a precise, sobering decomposition of a claim the field makes loosely: that living tissue's spontaneous activity is high-quality randomness. The decomposition says: irregular spiking dynamics are necessary but nowhere near sufficient; the readout transform and post-processing do decisive work that biology does not do for you. A dish of cortical organoids exhibits asynchronous, Poisson-like, chaotic-looking firing that is simultaneously its computational substrate. That is exactly the regime this paper shows can yield passable random bits, so an organoid computer could in principle mint entropy from the same activity it uses for reservoir computation, saving interface energy and hardware.

But the same coincidence creates a security problem nobody in biological computing is modeling yet. The authors note in passing that the balanced chaotic state is simultaneously a reservoir: the identical dynamics can be stabilized into a computer or read as a random source. In an organoid, these are not two devices, they are one. If you harvest randomness from your computing tissue, your output bits are a many-to-one projection of the internal state that carries your computation. A stream you publish or reuse as key material leaks, in principle, information about the trajectory you were using to compute. NIST will never flag that, because the bits can be perfectly random and still correlated with your confidential computation. Any real wetware random-number service needs an information-flow wall between the computation and the entropy tap, plus the whitening this paper shows is mandatory.

The threat has a mirror image: if spontaneous organoid activity is ever marketed as a true random number generator, this paper is the benchmark it must clear. Raw spikes fail; refractory structure is detectable; only engineered transforms pass. And the dual-use framing cuts in: a chaotic living network with 2-to-the-power-N-squared instance diversity and NFC-level power draw is a tantalizing description for someone building low-detectability key generators, a fact regulators of biological computing should read before assuming wetware randomness is a benign application. The realistic opportunity is narrower and more useful: treat entropy harvesting as an engineering discipline with per-instance QC, conditioning, and explicit separation from computation, exactly as this paper does for silicon.

The bottom line

Established: a binarized, quantized balanced spiking network in the spike-chaos regime, read out through a dynamic lookup table and two passes of XOR whitening, produces bit streams that match established generators on NIST SP-800-22, TestU01 SmallCrush, and whitened Dieharder, and does so at 5.45 mW and 120 kbps in fabricated FPGA hardware. Hypothesis: that the same architecture ported to ASICs or neuromorphic processors keeps sub-milliwatt power at higher bandwidth, and that the security story (state secrecy, side channels) holds up under formal analysis. What would confirm it: independent cryptanalysis and published cycle-time bounds. What would break it: a state-reconstruction attack from observed bits, or evidence that quality varies more sharply with instance choice than reported. For organoid intelligence the lesson stands regardless: biological noise is neither free randomness nor pure nuisance; it is a conditioned, audited resource that must be firewalled from the computation it shares a substrate with.