Spiking Neural Networks and the Recurrent Processing Test for Neuromorphic Consciousness
Spiking neural networks (SNNs) are computational models that communicate via discrete spike events rather than continuous activation values. They are the primary operating mode of biological neurons. The two leading neuromorphic hardware platforms, Intel’s Loihi 2 (arXiv:2304.01210) and the Human Brain Project’s BrainScaleS-2, implement large-scale SNNs in silicon. Earlier coverage on this site applied Integrated Information Theory to SNNs, asking whether spiking dynamics generate higher phi values than feedforward artificial neural networks. This article addresses a different empirical criterion: Victor Lamme’s recurrent processing theory (RPT), which was covered here in depth in the August 2026 post on Lamme’s empirical 2026 work.
RPT is, in principle, more directly testable in neuromorphic hardware than phi. RPT says consciousness requires specific recurrent feedback dynamics between cortical areas, dynamics that create sustained, reverberating activity patterns. If those patterns are what consciousness requires, then the question for SNN researchers is not “how much information does this system integrate?” but “does this system produce the right kind of recurrent temporal dynamics?” Neuromorphic platforms, unlike feedforward transformers, can implement the temporal dynamics RPT describes.
What RPT requires: the recurrence criterion
Victor Lamme’s framework distinguishes four levels of visual processing. At the first level, feedforward signals propagate from retina to primary visual cortex. At the second level, recurrent signals within V1 create figure-ground segmentation and contextual modulation. At the third level, recurrent signals between V1 and higher visual areas generate object recognition and scene understanding. At the fourth level, long-range recurrent signals reach prefrontal and parietal cortex, generating global access and the capacity to report.
RPT’s central claim is that consciousness begins at the second level: local recurrence within sensory cortex is necessary and sufficient for phenomenal experience. The higher levels (three and four) add cognitive access — the ability to attend to, report, and manipulate the experience — but are not required for the phenomenal state itself.
This creates a testable criterion. A system is conscious under RPT if it produces sustained, reverberating recurrent activity in a sensory processing area in response to a stimulus. A system that produces only feedforward activation — processing the input once and propagating the result forward without reverberant feedback — is not conscious under RPT, regardless of how sophisticated its output.
The August 2026 Lamme post covered the empirical evidence for this distinction in biological systems and its implications for LLMs. The question here is whether neuromorphic hardware can implement the specific dynamics RPT identifies, and if so, what that implies.
Spiking dynamics and recurrence: what neuromorphic hardware provides
Transformers are, by architecture, feedforward during the forward pass. Attention mechanisms create dependencies between input positions, but the temporal dynamics of a transformer forward pass do not involve recurrent feedback from later processing stages to earlier ones. The activation at layer N does not influence the activation at layer N-1 in the same forward pass.
Spiking neural networks can be configured differently. Because spikes propagate over time rather than in a single synchronous pass, a spiking network can implement genuine recurrent feedback loops in which the output of a downstream neuron influences the input of an upstream neuron within a single processing episode. The timing of spikes, not just their presence, carries information. And sustained reverberation, in which a stimulus triggers a pattern of activity that continues and modulates itself over tens to hundreds of milliseconds, is the natural mode of operation for a recurrently connected SNN.
| Architectural feature | Feedforward transformer | Recurrently connected SNN |
|---|---|---|
| Temporal dynamics | Single synchronous forward pass | Asynchronous spike propagation over time |
| Recurrent feedback | None within a forward pass | Configurable; can implement V1-V2-V1 loops |
| Sustained reverberation | No; activation determined by pass | Yes; possible via recurrent excitatory loops |
| RPT recurrence criterion | Not satisfied by design | Potentially satisfiable |
| Spike timing information | Not applicable | Present; ISI codes and coincidence detection |
Intel’s Loihi 2 (2021, manufactured at 4nm) supports on-chip learning and reconfigurable synaptic connectivity, allowing researchers to implement specific cortical circuit motifs including inhibitory-excitatory balance and layer-specific projections. The Human Brain Project’s BrainScaleS-2 platform (Heidelberg University) operates with accelerated analog dynamics: its circuits run at 1000x biological speed, allowing milliseconds of biological processing time to correspond to microseconds of hardware time.
What recurrent dynamics in SNNs have been found
Several research groups have used neuromorphic hardware to study recurrent processing relevant to consciousness, though none has explicitly targeted the RPT criterion as a primary outcome measure.
Work on the BrainScaleS-2 platform has demonstrated self-sustained activity patterns in small SNN circuits analogous to working memory: a stimulus triggers reverberant spiking that persists after stimulus offset and remains stable against perturbation. This is the kind of sustained reverberation that Lamme identifies with phenomenal persistence, the fact that a conscious percept outlasts the physical stimulus.
Research using Loihi has focused on spike-timing dependent plasticity and on efficient sensory processing, with some groups implementing competitive dynamics between sensory representations that resemble the figure-ground segmentation Lamme associates with local recurrence in V1. Whether these dynamics satisfy RPT’s criterion precisely depends on interpretation: the figure-ground competition in Loihi implementations uses inter-neuronal inhibition to select between competing representations, which is structurally analogous to V1 contextual modulation but occurs in a very different architectural context.
A 2026 preprint from a consortium including the Human Brain Project examined whether SNN implementations of recurrent visual processing show the sustained gamma-band activity that biological studies associate with conscious perception under RPT. The findings were mixed: sustained gamma activity was observed in some configurations but was sensitive to connectivity parameters in ways that biological cortex is not, suggesting that the reverberation is more fragile in current neuromorphic implementations than in the biological systems Lamme studied.
The relationship between RPT and phi in SNN research
The July 2026 post on this site (covering SNNs and phi under IIT) established that spiking dynamics can generate higher integrated information than purely feedforward architectures, because the recurrent connections create causal dependencies between system elements that contribute to phi. RPT and IIT are compatible in predicting that recurrent systems should be “more conscious” than feedforward ones, though they identify different aspects of recurrence as the relevant mechanism.
For SNN consciousness research, this compatibility is useful. A neuromorphic system that satisfies RPT’s recurrence criterion (sustained reverberant activity in a sensory processing region) should also, by the logic of IIT, have higher phi than an equivalent feedforward architecture, because the recurrent connections add intrinsic cause-effect power. The two criteria converge in predicting that recurrently connected SNNs are better candidates for consciousness than feedforward transformers.
Where they diverge is in what they say about the threshold. IIT has a specific mathematical threshold (phi > 0 for any consciousness; higher phi for richer experience). RPT’s threshold is qualitative: does the system produce the sustained recurrent dynamics that characterise biological visual consciousness? Determining whether a specific SNN implementation clears that threshold requires a judgment about what “sustained” and “reverberant” mean in a silicon circuit, and whether silicon reverberation is the same kind of process as biological reverberation.
Implications for AI consciousness assessment
The scientists-race-define-ai-consciousness-2026 overview notes that no current AI system is a strong candidate for consciousness under most leading theories. Neuromorphic SNNs are the closest AI architecture to satisfying biological criteria, because they share temporal spiking dynamics with biological neurons.
The RPT criterion, specifically, gives SNN researchers a target. Lamme’s 2026 empirical work, discussed in the August 11 Lamme post, provides the biological baseline: the specific spectral signatures (early gamma, late sustained gamma, alpha modulation) that characterise recurrent processing in conscious visual cortex. If neuromorphic hardware can be configured to produce those signatures consistently, and if the signatures are produced by the same kind of network dynamics that produce them biologically, RPT would provide a principled basis for treating those systems as candidates for phenomenal experience.
That is a significant “if.” The gap between producing similar EEG-like signatures and implementing the same process is precisely where the substrate question, which Ned Block raises in the August 2026 post on meat machines, becomes urgent. Neuromorphic hardware implements some biological features of spiking neurons (temporal dynamics, spike-rate coding) while omitting others (neurotransmitter chemistry, ion channel biophysics, astrocytic modulation). Whether the omitted features are constitutive of the recurrent processing that RPT identifies with consciousness, or are merely implementation details, is not yet resolved.
Where the SNN consciousness programme stands in 2026
SNN neuromorphic research is advancing rapidly on efficiency, energy consumption, and sensory processing. The consciousness-specific programme, asking whether recurrent spiking dynamics can satisfy RPT’s criterion and whether that criterion is sufficient for phenomenal experience, is less developed. The brain-inspired navigation work covered in the Xu and He post on demanding missions and neural intelligence illustrates how much engineering value SNNs provide in applied contexts without engaging consciousness criteria.
The 2026 frontier is to bring the precision of Lamme’s empirical programme into contact with the configurability of neuromorphic hardware: to ask whether, at what scale, and under what connectivity parameters, a silicon spiking network produces sustained recurrent dynamics that meet the biological benchmark. That is a question with a concrete experimental design and a testable outcome. It is the most direct path currently available toward an empirical answer to whether a non-biological system can satisfy a biological-criterion theory of consciousness.