Neural circuits improve the fidelity of a representation not by shouting louder but by going quieter: inhibitory neurons suppress untuned background activity and responses to non-preferred features, lowering the noise floor so that the target signal stands out. This "inhibitory sharpening" has been demonstrated at every level of the nervous system, from cricket wind-detection interneurons to human primary auditory cortex, and it has a striking non-invasive analog: applying 5 Hz repetitive transcranial magnetic stimulation over visual cortex improves brain-computer interface decoding by suppressing background brain activity rather than amplifying the evoked signal (Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention). But that stimulation was delivered blind to the brain's own rhythm — an unphased, open-loop train that strikes the ongoing alpha oscillation at random moments, sometimes amplifying and sometimes cancelling.
Meanwhile, a separate literature shows that the effect of any neuromodulatory pulse depends strongly on the millisecond-scale phase of the ongoing rhythm it lands on: near-threshold visual stimuli are perceived or missed depending on pre-stimulus alpha phase (The Phase of Ongoing EEG Oscillations Predicts Visual Perception), and single TMS pulses over visual cortex change phosphene detection probability by roughly 15% between opposite phases of the 10 Hz rhythm (The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception). The proposal here is to marry these two findings: to build a closed-loop system that tracks the instantaneous phase of the endogenous alpha rhythm over visual cortex and delivers suppression-seeking pulses at a specific, predicted phase angle, testing the hypothesis that background noise quenching — the mechanism behind inhibitory sharpening — is phase-gated and can be deliberately maximized.
Core Claim
The central hypothesis is that non-invasive stimulation suppresses ongoing cortical noise most effectively when pulses are delivered at a particular phase of the local alpha rhythm, because that phase is when inhibitory interneuron pools are most recruitable. Three observations motivate this. First, optogenetic activation of parvalbumin interneurons in mouse V1 sharpens visual discrimination only within a strict early time window, and impairs it outside that window (Context-dependent activation of V1 parvalbumin interneurons enhances visual discrimination) — timing determines whether inhibition helps or hurts. Second, that window was defined relative to external stimulus onset, not to the endogenous oscillation; yet in silico, interneuron-mediated gamma rhythms create intrinsic cycles of excitability that gate task output (Neural oscillation in low-rank SNNs: bridging network dynamics and cognitive function). Third, TMS pulses at sufficient intensity delivered near the peak of a rhythm can trigger a phase inversion, plausibly by recruiting previously silent, higher-threshold neuronal pools (TMS-induced phase resets depend on TMS intensity and EEG phase) — a candidate mechanism for selectively recruiting inhibitory cells, which typically have higher thresholds, at a chosen phase.
A secondary claim addresses a real tension in the modeling literature. Theory validated in Parkinsonian rats shows that networks at low baseline synchrony (5–10% of saturation) are mathematically biased toward amplification, not suppression, under phase-locked stimulation (Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation). Yet task-engaged cortex needs exactly the opposite — variability quenching toward a stable fixed point, as demonstrated across human fMRI and primate spiking recordings (Task-evoked activity quenches neural correlations and variability across cortical areas). Our resolution: the amplification bias applies to pulse trains attempting sustained desynchronization of a population oscillator. Quenching does not require desynchronizing the rhythm; it requires recruiting inhibition at the right phase, even intermittently. We will test whether sparse, phase-locked pulses — far below the dose of continuous open-loop trains — can quench background variability in a low-synchrony network.
Research Program
The program has three parts, moving from simulation to a closed-loop human experiment.
In the modeling part, we extend the low-rank spiking network of Neural oscillation in low-rank SNNs: bridging network dynamics and cognitive function, which already generates endogenous oscillations with Dale's-principle connectivity and shows phase-dependent processing, by adding a distinct inhibitory population with a nonlinear activation threshold (motivated by the competitive pyramidal-interneuron interactions used in Context-dependent activation of V1 parvalbumin interneurons enhances visual discrimination). We simulate phase-locked pulses across the full cycle and measure a phase-conditioned quenching index,
$$Q(\theta) = 1 - \frac{\mathrm{Var}(y \mid \theta)}{\mathrm{Var}(y)},$$
where $y$ is population activity, $\theta$ is the pulse phase angle, and $\mathrm{Var}(y)$ is background variability without stimulation. The model predicts the suppression-optimal phase $\theta^*$ and its dependence on pulse intensity, which we can cross-check against the empirical finding that phase-dependent modulation emerges above roughly 40% of resting motor threshold with an optimal window near 70–80% (TMS-induced phase resets depend on TMS intensity and EEG phase). We also verify the model against the amplitude/phase-response relationship validated in the Parkinsonian rat work, where the amplitude response curve correlated negatively with the derivative of the phase response curve at $R = -0.84$ across 12 of 13 animals (Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation) — if our extended model violates this relationship, the inhibitory recruitment mechanism is mis-specified.
In the human experiment, we recruit roughly 30 healthy subjects performing a steady-state visual evoked potential (SSVEP) brain-computer interface task — a paradigm in which flickering on-screen targets drive narrow-band EEG responses that a decoder classifies. Real-time phase of each subject's individual alpha rhythm over occipital cortex is tracked with the endpoint-corrected Hilbert transform, which reduces endpoint phase estimation error from approximately $47^\circ$ to $9^\circ$ compared to the standard Hilbert transform (Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence). Single TMS pulses at 70–80% of resting motor threshold are delivered at four phase bins (peak, rising flank, trough, falling flank), plus a dose-matched open-loop train replicating the 5 Hz protocol of Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention, plus sham. We follow that paper's design of four SSVEP frequency bands (8–12, 18–22, 28–32, 38–42 Hz) and its Pre/Post0/Post20 measurement schedule, and we adopt its decomposition logic in reverse: rather than asking whether signal power grew, we ask whether background variance shrank, computing the same style of variability-quenching statistic established in Task-evoked activity quenches neural correlations and variability across cortical areas.
In the benchmark part, the decoding pipeline is fixed before any stimulation data are collected, using the BETA database — 64-channel EEG from 70 subjects on a 40-target cued-spelling SSVEP task (BETA: A Large Benchmark Database Toward SSVEP-BCI Application) — to select and tune the supervised decoder (msTRCA, which the BETA paper reports as generally outperforming other supervised methods at shorter data lengths). The wearable 12-target SSVEP dataset with 102 subjects and 240 epochs per subject (An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces) then tests robustness on low-channel, high-impedance hardware, since a practical closed-loop system cannot assume a 64-channel cap. A final component addresses the field's open phase-drift problem: the optimal phase is expected to drift within a session, so the controller adapts $\theta^*$ online, and we test whether baseline rhythm regularity predicts response, as it did with 97% accuracy for tremor responders in Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence.
Baselines
The naive reference point is no stimulation: the same SSVEP task decoded by the same frozen pipeline with sham stimulation. On the wearable benchmark, this corresponds to published offline accuracy of $0.93 \pm 0.01$ (wet electrodes, FBTRCA, 2-second data length) and information transfer rate of $92.71 \pm 2.06$ bits/min, against $0.73 \pm 0.03$ and $64.83 \pm 3.74$ bits/min for dry electrodes (An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces); for the BETA database itself, no single-number decoding accuracy is reported on Lacuna, so our own sham arm supplies that reference.
The strongest existing method at this intersection is open-loop 5 Hz rTMS over V1, which improved SSVEP decoding accuracy by +7.53% in the 18–22 Hz band and +11.4% in the 28–32 Hz band relative to sham, with gains persisting at least 20 minutes, via suppression of background activity (Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention). We will run this protocol ourselves under matched conditions — same subjects, same task, same session schedule, and, critically, the same total pulse count as our phase-locked arm — because the honest competition is not "phase-locking versus nothing" but "phase-locking versus the same physical dose delivered phase-blind." Beating this baseline means achieving a larger accuracy gain than +11.4% in the high band at equal or lower pulse count, or matching the gain with a substantially smaller dose.
Two further comparisons isolate the proposal's key ingredients. A random-phase arm uses the identical closed-loop hardware, intensity, and pulse count but draws the target phase uniformly at random; it separates the effect of phase-locking from the effect of delivering any pulsed stimulation, since if pulse timing is irrelevant, random-phase should match $\theta^*$-locked delivery. And an algorithmic competitor applies state-dependent decoding — the approach that improved extracted information by 21.74% in a bidirectional rat BMI by factoring pre-stimulus neural state into the decoder (State-Dependent Decoding Algorithms Improve the Performance of a Bidirectional BMI in Anesthetized Rats) — to our sham data. This comparison is cross-domain (their numbers come from intracortical rat data, not human EEG) so it is reported as context, not as a head-to-head threshold, but the decisive question it poses is whether biophysical state control adds anything beyond the best statistical discounting of the same state.
Why not just…?
Why not just use open-loop repetitive stimulation? This is the most serious objection, because the open-loop protocol already works: +11.4% decoding gain with a standard clinical stimulator and no real-time hardware (Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention). The counter-evidence is that unphased pulses land at random phases, so amplifying and suppressing pulses average out and the effective dose is wasted; in essential tremor, non-phase-locked cerebellar stimulation tested as a control was ineffective, while phase-locked stimulation at the optimal lag reduced tremor amplitude by 21.5% (Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence). Our dose-matched open-loop and random-phase arms settle this directly and honestly: if phase-locked pulses at the model-predicted $\theta^*$ do not outperform dose-matched open-loop trains, the phase-gating hypothesis is wrong and the simpler clinical protocol should be preferred.
Why not just improve the decoding algorithm? Bigger decoders are cheaper than stimulators. But the BETA database shows that when sensor-level SNR drops in an unshielded environment, all eleven state-of-the-art supervised and training-free algorithms degrade together — post-processing cannot manufacture information the sensors did not capture (BETA: A Large Benchmark Database Toward SSVEP-BCI Application) — and the rTMS paper's authors make the same point explicitly: SNR, not decoding, is the binding constraint (Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention). The passive state-discounting route is also limited: the time-averaged state decoder in the rat BMI was substantially inferior to the one capturing full oscillatory dynamics (State-Dependent Decoding Algorithms Improve the Performance of a Bidirectional BMI in Anesthetized Rats), and perception varies ~15% across phase angles independently of slow power shifts (The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception). Still, we treat the algorithmic route as a real threat: our design runs the strongest decoder on both sham and stimulated data, so if the decoder alone closes the gap that stimulation closes, stimulation loses. The converse test — whether phase-locked quenching plus decoding beats decoding alone — is where the proposal's value would be established.
Why not just lock the pulses to stimulus onset instead of to oscillation phase? This is cheaper than real-time phase tracking and already has empirical support: PV activation sharpened discrimination only within the first 120 ms after stimulus onset (Context-dependent activation of V1 parvalbumin interneurons enhances visual discrimination). The two timings are, however, dissociable, because stimulus onset lands at a different oscillation phase on each trial. Our design deliberately jitters stimulus onset relative to ongoing phase, so onset-locked and phase-locked accounts make different predictions: if the sharpening window is genuinely phase-gated, then pulses at $\theta^*$ should help regardless of onset timing, and onset-locked pulses should help only on trials where onset happened to fall in the receptive phase. If instead the onset-locked account fully explains the effect, the phase term in our regression will be null — and that would itself be an informative answer to the open question left by the mouse optogenetics work.
Proposed Experiments and Decisive Checks
The decisive test is phase-specificity: decoding accuracy and the quenching index $Q$ should vary systematically and sinusoidally with stimulation phase, with a maximum at the model-predicted $\theta^*$ and a minimum roughly $\pi$ away, mirroring the antiphase structure seen in attention-modulated reaction times at frontal 3 Hz (Alerting attention is sufficient to induce a phase-dependent behavior that can be predicted by frontal EEG). A flat phase profile would refute the core claim even if overall accuracy improved. Second, the mechanism check: any accuracy gain must decompose into reduced background variance with unchanged evoked-signal amplitude — the signature established by the rTMS study (Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention) — and, following Task-evoked activity quenches neural correlations and variability across cortical areas, into reduced inter-channel noise correlations and increased dimensionality of the evoked response. Third, the dose-efficiency check: phase-locked pulses should reach the open-loop gain (+11.4% in the high band) with a fraction of the pulse count; energy-per-decision matters if such a loop is ever to run on the implantable hardware described in Intelligent Neural Interfaces: An Emerging Era in Neurotechnology, where on-chip classification already runs at 0.227 µJ per decision. Fourth, the drift check: the adaptive controller should maintain gains across a 20-minute post-stimulation window, and baseline rhythm regularity should predict responders. If the phase-gating hypothesis survives these checks, the payoff is a concrete design rule — suppression is a phase to be aimed at, not a train to be blasted — and a bridge between the cellular biology of inhibitory sharpening and the engineering of closed-loop neuromodulation.
References
- Enhancing visual brain-computer interface through V1-targeted RTMS by modulating visual attention
- The Phase of Ongoing EEG Oscillations Predicts Visual Perception
- The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception
- Context-dependent activation of V1 parvalbumin interneurons enhances visual discrimination
- Neural oscillation in low-rank SNNs: bridging network dynamics and cognitive function
- TMS-induced phase resets depend on TMS intensity and EEG phase
- Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation
- Task-evoked activity quenches neural correlations and variability across cortical areas
- Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence
- BETA: A Large Benchmark Database Toward SSVEP-BCI Application
- An Open Dataset for Wearable SSVEP-Based Brain-Computer Interfaces
- State-Dependent Decoding Algorithms Improve the Performance of a Bidirectional BMI in Anesthetized Rats
- Alerting attention is sufficient to induce a phase-dependent behavior that can be predicted by frontal EEG
- Intelligent Neural Interfaces: An Emerging Era in Neurotechnology