Source-linked AI summary
Simulating Synchrony Loop Networks in the Open Source RISP Neuroprocessor
Jackson Mowry, Patrick Abbs
TL;DR
SNNs offer efficient event-driven computation but still lag on complex learning tasks, and standard neuromorphic architectures do not readily support SLP mechanisms. This paper implements SLP on the RISP neuroprocessor and evaluates it on unsupervised musical-instrument clustering. The system achieves clustering performance comparable to DBSCAN while improving runtime speed by over 50× and enabling real-time processing.
Problem
Method
Results
Takeaways & Limitations
The results demonstrate the feasibility of translating SLP mechanisms into neuromorphic systems and show that population-level learning dynamics can be explicitly engineered with localized signaling.
Source: §6 · p. 8The evaluation uses a relatively constrained instrument-clustering task, and RISP’s software support for SLP mechanisms is not yet translatable into hardware.
Source: §6 · p. 8
Abstract
from arXiv · showhide
Neuromorphic spiking neural networks (SNNs) offer a promising alternative to conventional deep neural networks for tasks with computational resource or data constraints. However, their practical applications have been limited by comparatively weak performance on complex learning tasks. Experimental approaches such as Synchrony Loop Propagation (SLP) increasingly seek to address this problem through more sophisticated and heterogeneous neuron models, and have achieved encouraging initial results. However, these neuron models do not readily translate to standard neuromorphic systems designed to support simple leaky integrate-and-fire neurons. We present an implementation of an SLP network on the RISP neuroprocessor, an event-driven neuromorphic simulation platform, and evaluate its performance on an unsupervised musical instrument clustering task. The network achieves clustering accuracy comparable to the DBSCAN algorithm, while providing over an order of magnitude improvement in runtime speed and computational efficiency relative to a prior non-neuromorphic SLP implementation. These results demonstrate that SLP's core mechanisms can be effectively translated into a neuromorphic architecture to support complex unsupervised learning. More broadly, this work highlights the potential of heterogeneous and extensible neuron models to expand the design space of neuromorphic systems to more complex learning tasks.
1. Introduction
The introduction motivates richer, deliberately coordinated SNN learning and presents an SLP implementation on RISP evaluated for musical-instrument clustering. The reported system matches DBSCAN clustering performance and exceeds real-time processing speed.
1. Introduction
- SNNs offer distributed, event-driven computation, but lag DNNs on complex learning tasks; limited neuron heterogeneity and indirectly arising population dynamics are proposed as contributing factors.Most SNN frameworks primarily rely on forward spikes and long-term weight updates.Source: §1 · p. 1
- SLP combines heterogeneous neuron roles with continuous unsupervised bidirectional feedback loops to produce coordinated population-level learning under localized signaling constraints.Prior work reported strong performance on blind sound source separation (BSS).Source: §1 · p. 1
- Conventional neuromorphic systems do not fully support SLP mechanisms, and early non-neuromorphic implementations incurred overhead and latency that made real-time edge deployment unsuitable.The paper extends RISP, originally designed for homogeneous integrate-and-fire neurons, to support more complex neuron and synapse models, then evaluates translated SLP layers as a standalone sound-profile clustering network.Source: §1 · pp. 1–2
- Over 50× runtime speed improvement and clustering performance comparable to DBSCAN are reported for the RISP implementation, which surpasses real-time processing.The results also report support for all translated SLP mechanisms.Source: §1 · p. 2
2. Related Work
Prior SNN clustering work uses local competition and STDP, but the described training setup assumes advance knowledge of the number of cluster centers and uses cluster-specific output information.
2. Related Work
- STDP-based learning with lateral inhibition has been used to cluster arbitrary data, with local competition forming more separated cluster centers.The described training requires knowing the number of centers in advance and sends spikes to the “correct” output neuron for each cluster, information unavailable in a truly unsupervised setting.Source: §2 · p. 2
- The prior method requires knowing ahead of time how many cluster centers should be fit.Source: §2 · p. 2
- Its learning rule sends spikes to the “correct” output neuron for each cluster, requiring information unavailable in a truly unsupervised environment.Source: §2 · p. 2
3. Synchrony Loop Propagation
Synchrony Loop Propagation (SLP) is an unsupervised, spike-driven learning framework in which local synaptic and neuronal interactions support coordinated population-level learning dynamics.
3. Synchrony Loop Propagation
- The section introduces SLP’s core mechanisms and explains how localized interactions give rise to emergent population-level learning dynamics.Exact implementation equations are provided in Section 4.Source: §3 · p. 2
- SLP uses temporal spikes to drive synaptic plasticity, while synaptic information also modifies presynaptic output routing rather than only postsynaptic input integration.Source: §3 · p. 2
- This coupling creates continuous bidirectional feedback through which local interactions propagate across the network and produce coordinated global behaviors learning from unstructured sensory inputs.Source: §3 · p. 2
3.1. Postsynaptic Prediction
SLP characterizes learning through interactions across a population: synapses encode expected input structure and timing, and pressure dynamics indicate how well a neuron predicts its inputs.
3.1. Postsynaptic Prediction
- SLP learning cannot be characterized at the level of individual neurons in isolation; postsynaptic and presynaptic processes must be considered together.Source: §3.1 · p. 2
- Unlike conventional models, SLP input synapses encode temporal predictions of input signals without directly modifying those signals.In conventional neuron models, synaptic weights scale incoming signals.Source: §3.1 · p. 2
- Each synapse has a long-term strength and short-term pressure variable; in this formulation, strengths define a normalized distribution of expected input spike rates.For strengths {10.0, 50.0, 40.0}, the expected relative spike distribution is {0.1, 0.5, 0.4}.Source: §3.1 · p. 2
- Relative pressure across input synapses predicts the source of the next spike, with greater pressure indicating a more likely source.For pressures {0.5, 0.9, 1.6}, synapse 3 is predicted as most likely and synapse 1 least likely to receive the next spike.Source: §3.1 · p. 2
- On an input spike, the neuron removes pressure equal to the spike magnitude from the receiving synapse and redistributes it across synapses in proportion to their strengths.This gives each synapse pressure inflow tied to its predicted relative spike rate and outflow tied to its actual relative spike rate.Source: §3.1 · p. 2
- Pressure dynamics encode prediction accuracy: when strengths match true input rates, pressure concentrates on the correct synapse immediately before each spike.Misaligned strengths degrade predictive alignment, making postsynaptic pressure a local measure of how well a neuron models its input.Source: §3.1 · p. 2
3.2. Presynaptic Routing
SLP completes postsynaptic pressure prediction with presynaptic routing that allocates spikes according to synaptic pressure. This bidirectional feedback favors neurons that predict spike timing and concentrates activity into representations of input structure.
- Presynaptic routing completes postsynaptic pressure prediction by directing spikes toward neurons that accurately predict their timing.Source: §3.2 · pp. 2–3
- Presynaptic neurons distribute each firing spike across output synapses in proportion to pressure, favoring postsynaptic neurons that better predict the spike.The basic formulation normalizes total spike magnitude to 1.0.Source: §3.2 · p. 3, Figure 1 · p. 3
- Predictive neurons attract larger spike portions that further amplify pressure redistribution, while poorer predictors attract smaller portions that depress it.Together, the decentralized feedback loop concentrates stimulus-driven activity into neurons that represent the input structure.Source: §3.2 · p. 3
3.3. Emergent Learning
SLP uses pressure-driven routing to localize synaptic learning, allowing neurons to specialize to recurring input patterns. Different stimuli can recruit different neurons, and instrument clustering uses learned synaptic strengths to represent each instrument’s temporal spike distribution.
- Each postsynaptic input spike increases the corresponding synaptic strength in proportion to the received signal’s magnitude.Source: §3.3 · p. 3
- Pressure-driven routing localizes updates to the neuron with the strongest predictive alignment, progressively adapting its strengths to the observed input pattern.Repeated exposure reinforces the alignment and specializes the neuron to represent that stimulus.Source: §3.3 · p. 3
- When a different stimulus appears, routing shifts toward a more predictive neuron; learned neurons can reactivate when their original stimulus returns.This produces competitive, unsupervised specialization across input patterns.Source: §3.3 · p. 3
- For instrument clustering, synaptic strengths encode each instrument’s temporal spike distribution and support online learning as stimuli are observed.Source: §3.3 · p. 3
3.4. SLP Extension and Modulation
The instrument-clustering SLP network extends its core mechanisms with specialized neuron behaviors. These extensions address fine temporal structure, stabilize learned representations, gate synaptic updates, and coordinate competition in the final clustering layer.
3.4.1. Continuous Signal Release.
- Continuous output signals address the poor speed and precision of discrete spikes for representing fine-grained temporal structure such as sub-millisecond auditory features.The auditory input layer communicates frequency components downstream using continuous real-time signaling.Source: §3.4.1 · pp. 3–4
3.4.2. Representational Maturity.
- A maturity variable scales down synaptic strength increments as total input strength approaches a fixed upper limit, shifting neurons from plasticity toward stable encoding.Learned representations are retained while remaining capable of limited refinement.Source: §3.4.2 · p. 4
3.4.3. Multiple Signal Types.
- Exploratory signals gate whether activating signals may modify long-term synaptic strength, protecting neurons that do not match a stimulus from updates.Some postsynaptic neurons accept activating signals only after exploratory input reaches a threshold.Source: §3.4.3 · p. 4
3.4.4. Neuron Subpopulation Coordination.
- A coordinator monitors exploratory accumulations and enables activating-signal absorption only for the strongest responder, adding winner-take-all competition in the final clustering layer.This makes stimulus activation mutually exclusive across neurons in that layer.Source: §3.4.4 · p. 4
4. RISP SLP Implementation
The section extends RISP beyond homogeneous LIF networks to implement a heterogeneous five-layer SLP network for standalone sound-profile clustering, testing support for SLP mechanisms under neuromorphic constraints and their efficiency benefits.
4.1. Frequency Neurons
- Frequency neurons emit graded spikes for spectral amplitudes, routing them by synaptic pressure toward pitch neurons that group frequencies from one sound source.Each neuron represents frequencies from 10 Hz to 24 kHz in 10 Hz increments.Source: §4.1 · p. 4
4.2. Pitch Neurons
- Pitch neurons detect configured harmonic series and use recency-weighted input and synaptic pressure feedback to consolidate routing to the most active matching pitch neuron.They pool input by harmonic and route each above-threshold pool to its corresponding harmonic neuron.Source: §4.2 · p. 5, Figure 2 · p. 5
4.3. Harmonic Neurons
- Harmonic neurons passively integrate signals into pitch-invariant harmonic profiles, which form the feature set learned by sound neurons.Their routing raises pressure to the fourth power and separates exploratory and activating signal copies.Source: §4.3 · pp. 5–6
4.4. Sound Neurons
- Sound neurons learn harmonic profiles as clusters, reactivating a previously learned representation when the same sound returns.Their pressure redistribution is scaled by maturity, weakening predictions in developing neurons and stabilizing them as they mature.Source: §4.4 · p. 6, §4.4.1 · p. 6
4.4.1. Incorporating Maturity.
- Maturity scales pressure redistribution to temper developing neurons’ predictions and scales strength increments inversely to slow learning as representations stabilize.This reduces incorrect pressure predictions during initial imprinting while supporting stronger predictions in mature neurons.Source: §4.4.1 · p. 6
4.4.2. Incorporating Split Signal Types.
- Separate exploratory and activating pathways let exploratory input affect pressure without changing synaptic strength, while activating signals are routed only to neurons with high exploratory input.The pathways use distinct pressure and signal processing; activating pressure is zero unless exploratory input exceeds a threshold.Source: §4.4.2 · p. 6
4.4.3. Incorporating Lateral Coordination.
- A coordinator gates activating reception to the sound neuron with the highest exploratory input, yielding one active neuron at a time as the cluster assignment.The coordinator sets activating pressure to 1 for the highest-input synapse and 0 for the others.Source: §4.4.3 · pp. 6–7, Figure 3 · p. 7
5. Results
The RISP SLP network was evaluated on unsupervised musical-instrument clustering, with DBSCAN as an accuracy baseline and the original SLP implementation as a speed baseline. SLP had the higher combined clustering score and ran faster than real time, although DBSCAN scored higher on clip-level NMI.
5. Results
- Evaluation streamed 100 tracks, each 23 seconds and containing five randomly ordered scales, derived from recordings of four musical instruments.The network produced a cluster assignment for each time step in audio streams containing unknown instrument mixtures.Source: §5 · p. 7
- Sound-layer neurons served as clusters, assigning each time point to a neuron with activating pressure and an open signal pathway.The network processed the frequency amplitudes from the transformed tracks as input.Source: §5 · p. 7
- Accuracy used clip-level NMI and within-clip primary-cluster Confidence; DBSCAN required no preset cluster count, and both methods were tuned on these metrics.NMI measured overlap with hidden clip labels; Confidence measured the average fraction of time steps assigned to each clip’s primary cluster.Source: §5 · p. 7
- DBSCAN scored NMI 0.79 versus SLP 0.63, while SLP scored Confidence 0.89 versus DBSCAN 0.45; their combined scores were 0.56 and 0.36, respectively.DBSCAN had higher clip-level NMI, while SLP had higher time-step Confidence.Source: §5 · p. 7, Figure 4 · p. 7
- 17.2 s per 23-second track on RISP versus 884.6 s for original SLP yielded a 51× speedup; RISP processing ran at 1.3× real time.The processing-time comparison used an identical subset of input audio tracks.Source: §5 · pp. 7–8
6. Conclusion
The RISP implementation preserves core SLP mechanisms and achieves performance comparable to DBSCAN on musical-instrument clustering while improving processing speed by over 50×. The conclusion frames this as evidence of feasibility, while noting task-scope and hardware-support limitations.
6. Conclusion
- RISP preserved SLP’s bidirectional, localized synaptic feedback and neuron heterogeneity, achieving performance comparable to DBSCAN on musical-instrument clustering and increasing processing speed by over 50×.The authors describe this as establishing the feasibility of translating SLP mechanisms into neuromorphic systems.Source: §6 · p. 8
- The findings suggest population-level learning dynamics can be explicitly engineered under localized signaling, with heterogeneous computation and communication tailored to tasks.The paper presents heterogeneity and engineered emergence as tools for task-adaptive network design, while calling for further investigation on complex real-world tasks.Source: §6 · p. 8
- The clustering task is constrained relative to real-world learning, which may involve more complex structure, greater noise, and feedback or supervision.The authors call for testing SLP on more challenging tasks to assess its broader applicability.Source: §6 · p. 8
- RISP’s software support for SLP mechanisms is not yet translatable into hardware, limiting the neuromorphic efficiency benefits available to SLP implementations.The authors identify hardware translation as a longer-term challenge because existing platforms often do not natively support features such as graded spikes and bidirectional synapses.Source: §6 · p. 8
- Future work proposes scaling SLP to blind sound source separation by expanding pitch models, adding source-tracking layers, and incorporating user feedback pathways.The target task involves learning, tracking, and separating concurrent sounds in dynamic auditory environments.Source: §6 · p. 8