Neuromorphic DBSCAN turns MEA output into clustered events

What the work claims

Rizzo and Plank at the University of Tennessee, Knoxville, extend their earlier neuromorphic implementations of DBSCAN into the time domain. DBSCAN classifies each point in a dataset as a core point of a cluster, a border point, or noise, using two parameters: ε, a radial distance, and minPoints, a density threshold. The new spatiotemporal variants add a temporal neighborhood ε_t, so events count as neighbors only if they fall within both a spatial radius and a time window. Two constructions are carried over from the prior work: a flat network that maximizes throughput and a systolic network that trades time for hardware¹.

This is a methods and characterization paper. The authors do not report clustering accuracy on real data, and they say so: comparing the spatiotemporal networks against the earlier spatial versions for efficacy, and deploying them on hardware to measure latency on live event-camera streams, are both named future work¹. What the paper does deliver is exact, constructive accounting: closed-form neuron and synapse equations, and software-constructed totals for two example design points, all released as open source.

How it works

The flat construction dedicates a column of five neurons to every pixel of the input grid, plus two extra neurons per pixel for each unit of temporal reach ε_t, giving RC(5 + 2ε_t) neurons for an R × C grid. Synapses carry events forward in time using delay lines, so older activity within the temporal neighborhood persists in the network as charge. Because the network is strictly feed-forward, it classifies every pixel of an input frame in a fixed five timesteps to solution and one timestep to reuse, regardless of grid size. The cost is wiring: for a 260 × 346 input with ε = ε_t = 4, the constructed network totals 1,169,480 neurons and 72,770,360 synapses, with a worst-case fan-in of 404¹.

The systolic construction routes events through a row-wise pipeline instead, shrinking the same 260 × 346 example to 24,180 neurons and 232,460 synapses, about a 48-fold reduction in synapses and a 48-fold reduction in neurons, at the price of 358 timesteps to solution for that example. The same point in the small 10 × 10 configuration, with ε = ε_t = 2, uses 330 neurons and 1,630 synapses and solves in 18 timesteps, versus 900 neurons and 12,316 synapses in five timesteps for the flat version. Maximum neuron thresholds are minPoints - 1 in every variant, and multi-edge synapse support would eliminate the memory neurons entirely, an optimization the authors quantify but could not implement because their RISP neuroprocessor does not support multi-edges¹.

For resource-constrained targets, the paper adds partial or segmented implementations: run the temporal DBSCAN on sub-regions of the field of view sequentially, trading latency for area in the same way the systolic construction trades throughput for space. The generated networks target RISP-class neuromorphic hardware and are portable to Loihi 2, microcontrollers, or FPGAs through the TENNLab software suite; the generator is public at the TENNLab-UTK repository¹.

Where a skeptic should push

The load-bearing assumption is that density-based clustering is the right semantics for event-sensor noise, and this paper does not test it. The motivation it cites is external: a prior 3D-augmented DBSCAN study showed that temporal DBSCAN denoising of event-camera datasets (N-Cars, ASL-DVS, N-MNIST, POKER-DVS) preserved classifier accuracy better than competing methods as injected noise increased. Those results used conventional computers, and transferring them to spiking hardware is precisely what remains undone here¹. The authors' own conclusion lists efficacy comparison and real-hardware latency measurement as future work, which is the honest way to frame it.

Second, the characterizations are synthetic constructions from the authors' generator, not silicon measurements; no chip area, energy per event, or clock frequency appears in the paper. Third, DBSCAN's noise class is only as good as the density model: ε and minPoints are global parameters, and event sources with spatially varying density break the single-ε assumption on any platform, neuromorphic or not.

Neuromorphic DBSCAN and the MEA event stream

The reason this paper belongs on an MEA site is structural, and it is the non-obvious part: a CMOS microelectrode array produces exactly the data type this network eats. A high-density array reports threshold crossings as timestamped events on a fixed electrode grid, which is an event-camera frame with the pixel lattice replaced by an electrode lattice. Every argument the authors make about event cameras, low size, weight, and power at the edge, constant-latency processing, and denoising before downstream inference, applies verbatim to the MEA back end, with electrode pitch substituted for pixel pitch and ε measured in electrode spacings rather than pixels.

The specific mechanism worth stealing is the fixed solution time. The flat network reaches a complete clustering verdict in five timesteps no matter how many electrodes produced events, because the computation is a feed-forward pass over pre-wired neighborhoods rather than an iterative search. For closed-loop MEA systems, where a stimulation decision must land within a bounded window after a burst onset, a clustering front end with an a priori latency bound is qualitatively different from a software DBSCAN whose runtime grows with event count. The systolic variant's 48-fold synapse reduction is the relevant trade for implantable or head-mounted front ends, where area and power, not latency, bind first.

The mapping of parameters is also instructive. minPoints translates directly to the minimum spike count that constitutes a burst, and the noise class is a built-in artifact rejector: stimulation artifacts and electrostatic pickup are precisely the high-amplitude, spatially incoherent events that appear as isolated density outliers. A segmented implementation, processing electrode tiles in turn, mirrors how production spike sorters already tile large arrays. And the multi-edge optimization the authors quantify but could not run is a concrete hardware requirement an MEA-specific neuromorphic chip could actually satisfy, since nothing in electrode physics forbids multi-edges.

The threats are equally concrete. A fan-in of 404 synapses per neuron in the large flat configuration exceeds the fan-in limits of current neuromorphic cores, so the constant-latency construction is not deployable as drawn; the systolic or segmented versions are the realistic paths, and both add latency or bookkeeping that the five-timestep headline hides. DBSCAN's global ε assumes uniform density, while cortical and organoid cultures have spatially structured firing that a single density radius will over-fragment in quiet regions and merge in hot ones. Most seriously, the efficacy gap is not a formality: event-camera noise is largely salt-and-pepper, but MEA artifact structure is correlated and physiological slow drift creates exactly the density nonstationarity DBSCAN handles worst. Until someone runs the spatiotemporal network on real MEA event streams and reports cluster purity against a reference sorter, this remains a blueprint with plausible physics, not a validated front end.

The bottom line

Established: DBSCAN with a temporal neighborhood can be constructed as a spiking neural network with exact resource counts, a five-timestep constant solution time in the flat variant, and open-source generators. Not yet established: that these networks cluster or denoise as well as software DBSCAN on any real dataset, spiking or conventional, since the authors defer efficacy and hardware measurement to future work. For MEA instrumentation, the value today is architectural: it demonstrates a fixed-latency, event-native clustering layer that maps one-to-one onto electrode-grid output, and it specifies precisely what a neuromorphic core would need (multi-edge synapses, fan-in headroom, or segmentation) to host it. The confirming experiment is straightforward to state: inject recorded organoid burst data through the network, compare its clusters and noise rejection against a reference spike-and-burst sorter, and report accuracy per unit power against a software baseline.