Closed-loop neuromodulation delivers electrical pulses timed to the phase of an ongoing brain rhythm — the position within the rhythm's cycle — because identical pulses have very different effects depending on when in the cycle they land. This logic has been validated causally: visual perception depends on the pre-stimulus alpha phase The Phase of Ongoing EEG Oscillations Predicts Visual Perception, identical TMS pulses to visual cortex produce different perceptual effects by phase The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception, and phase-locked cerebellar stimulation reduces essential tremor by a mean of 21.5% while non-phase-locked stimulation proves ineffective Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence. Yet the same pulse, delivered at the same phase, still produces wildly different outcomes from trial to trial, and the field has no account of why. The standard tool for mapping phase to effect — the Phase Response Curve (PRC), built by binning pulses into phase buckets and averaging the post-pulse responses across dozens of trials — is precisely the kind of cross-trial average that systems neuroscience has shown to be a statistical fiction elsewhere: averaging invented a "dual reach plan" that never existed on any single trial Single reach plans in dorsal premotor cortex during a two-target task and turned discrete decision "steps" into illusory smooth ramps Discrete stepping and nonlinear ramping dynamics underlie spiking responses of LIP neurons during decision-making.
The core proposal: replace the trial-averaged PRC with a single-trial, phase-modulated latent state-space model, and test the specific hypothesis that the trial-to-trial variability in phase-locked stimulation outcomes is explained by where the population sits on its low-dimensional manifold in the moments before the pulse — information that phase alone does not carry. If true, the field's central curve is an average over hidden states, and closed-loop controllers should be conditioned on latent state, not phase alone.
Core Claim
Prior work makes the pieces plausible but has never assembled them. Response variability is structured, not Poisson noise: subthreshold inputs are repeatable and the non-linear spike threshold amplifies microscopic fluctuations into macroscopic spike-count variance Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex, sensory responses flip deterministically across a bifurcation between activated and inactivated cortical states State-dependence of sensory-evoked responses in neocortex, and single-trial population dynamics follow orderly low-dimensional trajectories that predict behavior (reaction time rank correlation $r_s \approx -0.55$) An orderly single-trial organization of population dynamics in premotor cortex predicts behavioral variability. Separately, phase-locked stimulation modeling exists but only in trial-averaged, scalar form: the Kuramoto analysis of Parkinsonian rats fits pooled Amplitude and Phase Response Curves, validating the derivative relation with a mean correlation of R = -0.84, but constructs those curves by binning and averaging across trials, treating the network as a single phase and synchrony order parameter Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation. No study in the file has characterized the single-trial response of a population manifold to phase-locked pulses, and none has tested whether pre-pulse latent position explains outcome variability beyond phase. That is the gap this proposal occupies.
Formally, we propose a latent state-space model in which the instantaneous oscillatory phase $\phi_t$ modulates both the endogenous dynamics and the effect of a stimulation pulse $u_t$:
$$x_{t+\Delta} = A(\phi_t)\, x_t + B(\phi_t)\, u_t + \varepsilon_t, \qquad \varepsilon_t \sim \mathcal{N}(0, Q)$$
where $x_t$ is the low-dimensional latent population state, $A(\phi_t)$ is the phase-dependent autonomous vector field, $B(\phi_t)$ is the phase-dependent pulse effect, and $\varepsilon_t$ captures residual noise. The classical PRC is the conditional expectation of the response given phase alone, $\mathbb{E}[\text{response} \mid \phi]$ — an average over the latent positions $x$ that co-occurred with each phase bin. If $B$ genuinely varies with $x$ as well as $\phi$, that average is a hybrid curve that matches no single trial, exactly as the averaged PMd activity matched no single reach plan Single reach plans in dorsal premotor cortex during a two-target task. A practical constraint from the hardware literature shapes the design: stimulation artifacts saturate recording front-ends and blind sensors immediately after each pulse Intelligent Neural Interfaces: An Emerging Era in Neurotechnology, so the model must treat the post-pulse window as missing data and bridge it by prediction, and the latent filter must run causally — a unidirectional, streaming adaptation of machinery that LFADS-class models currently run offline over buffered multi-second trials Inferring single-trial neural population dynamics using sequential auto-encoders.
Research Program
The program has four components. First, we build the phase-modulated latent model above, starting from existing latent-dynamical-systems code LDNS codebase but replacing its fixed transition matrices with phase-parameterized ones and its bidirectional encoder with a causal filter. Second, we construct a simulation benchmark with known ground truth by extending the OpenAI Gym deep-brain-stimulation environment of coupled oscillator populations Reinforcement Learning Framework for Deep Brain Stimulation Study: populations of Bonhoeffer–van der Pol and Hindmarsh–Rose units, phase-locked pulse delivery, and — crucially, unlike that environment's scalar mean-field observations — full single-unit readout plus a heterogeneous latent structure in which a pulse causes a discrete latent jump only when the population is on one side of a state threshold, mimicking the bifurcation structure of Curto et al. State-dependence of sensory-evoked responses in neocortex and the threshold amplification of Carandini Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex. Third, we reanalyze real phase-locked stimulation data — the rat ECoG GPe-stimulation recordings of 13 Parkinsonian animals Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation and the human tremor kinematics during phase-locked tACS Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence — contingent on data sharing, with a rodent pilot as the fallback deliverable. Fourth, we close the loop: a controller conditioned jointly on estimated phase $\hat{\phi}_t$ and latent state $\hat{x}_t$, evaluated against phase-only control on the benchmark, which also lets us arbitrate the open contradiction in the literature about what closed-loop stimulation should even target — amplitude suppression of synchrony (the premise of the RL framework, with reported suppression coefficient $S \approx 33$ and equilibrium convergence within roughly 1,100 time steps Reinforcement Learning Framework for Deep Brain Stimulation Study) versus disruption of temporal coherence, which is how the tremor suppression of 21.5% actually worked Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence.
Baselines
The naive reference point is open-loop, fixed-rate stimulation with no phase or state tracking; the file reports that non-phase-locked stimulation proved ineffective for tremor Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence, so any closed-loop policy must beat it to justify its complexity. The strongest existing scientific competitor is the trial-averaged Kuramoto PRC/ARC analysis of Mirkhani et al. Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation, which achieved a mean correlation of R = -0.84 between the pooled Amplitude Response Curve and the derivative of the Phase Response Curve across 12 of 13 animals; we will run it on exactly the same pulse sets and data splits as our model, giving it the same trials, the same phase bins, and the same tuning budget, and the comparison metric will be held-out single-trial prediction of the post-pulse response (cross-validated log-likelihood and explained variance), a metric its pooled curves were never asked to meet. Beating it means predicting individual pulse outcomes significantly better than the phase-bin conditional mean. The strongest existing closed-loop system is the endpoint-corrected Hilbert transform controller, which reduced phase estimation error from approximately 47° to 9° and delivered the 21.5% mean tremor reduction Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence; in the closed-loop benchmark we will give the phase-only controller the same compute, pulse budget, and training episodes, and count a win only as greater disruption at equal pulses, or equal disruption at fewer pulses. Finally, two ablations isolate the key ingredient: the same latent model with $A$ and $B$ held constant in $\phi$ (testing whether phase modulation matters at all), and phase-blind latent models of the LFADS/GPFA family — where AutoLFADS scored 0.346 bits per spike versus GPFA's 0.187 on MC_Maze in the standardized benchmark Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity — which tests whether latent state alone, without phase, suffices.
Why not just…?
Why not just use GPFA or Gaussian smoothing to denoise single trials? Linear methods are the obvious cheap alternative, but the file's evidence is direct: on macaque reaching, LFADS predicted movement velocity at $R^2 = 0.91$ while Gaussian smoothing reached only 0.61 and GPFA 0.66, and on the Neural Latents Benchmark AutoLFADS scored 0.346 bps against GPFA's 0.187 Inferring single-trial neural population dynamics using sequential auto-encoders Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity. More pointedly for us, linear smoothing cannot represent discrete steps, and the single-trial structure we are hunting is exactly step-like: 28 of 40 LIP neurons favored stepping over ramping once spike history was modeled Discrete stepping and nonlinear ramping dynamics underlie spiking responses of LIP neurons during decision-making. We still run GPFA as a baseline rather than argue around it.
Why not just track one electrode's phase with the endpoint-corrected Hilbert transform? This is the real threat, and we treat it as the decisive test rather than a rebuttal. The phase-only approach already achieved 21.5% tremor suppression and predicted treatment responders at 97% accuracy from baseline tremor regularity Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence. But its own authors report that the optimal stimulation phase drifts over time and their static-lag algorithm cannot adapt, and the Kuramoto work concedes that recording far from the stimulation site obscured local dynamics Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation. The settling experiment is a pre-registered variance decomposition: on held-out trials, we compare how much predictable outcome variance phase alone explains versus phase plus latent state. If phase alone explains nearly everything, the phase-only program is vindicated and our claim fails — we commit to reporting that.
Why not just collect more trials and average harder? Averaging reduces noise variance at rate $O(1/n)$, but it does not shrink state-mixture bias: averaging across trials in which the population held different discrete states produces a hybrid curve that converges to the wrong answer, not the right one. That is precisely what Dekleva et al. showed — trial-averaged PMd activity implied a dual reach plan (~85% on their Dual Representation Index) while single trials held one plan at a time (3–6%), with the initial single-trial plan predicting the eventual choice about 75% of the time Single reach plans in dorsal premotor cortex during a two-target task. Our simulation experiment makes this quantitative for stimulation specifically: we can measure whether the pooled ARC converges to the true conditional response or to a hybrid as trial count grows, with ground truth known.
Proposed Experiments
Experiment 1 is the ground-truth simulation. In the extended gym, we generate populations whose pulses cause discrete latent jumps conditional on latent position, deliver matched pulse sets across the phase cycle, and ask three questions: whether the trial-averaged ARC reproduces a smooth curve — including the ARC–dPRC derivative relation at the Mirkhani-reported R = -0.84 level — even when no single trial follows that curve (a diagnostic warning that the pooled metric is insensitive to single-trial structure); what fraction of individual pulse outcomes the phase-bin conditional mean mispredicts; and how much held-out single-trial prediction improves when $\hat{x}_t$ is added to $\phi_t$ as a predictor. Experiment 2 moves to real data: fitting the phase-modulated model to the rat ECoG GPe-stimulation recordings Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation and the human tremor recordings Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence, with the primary metric being held-out single-trial post-pulse response prediction (phase-only versus phase-plus-latent), and a secondary analysis testing whether the per-animal gain from adding latent state scales with the trial-to-trial variability that the 5–10% synchrony regime of Mirkhani et al. implies should be substantial. The file reports no result on Lacuna for single-trial stimulation-outcome prediction on any of these datasets, which is itself the gap; if the recordings cannot be shared, the deliverable shifts to a rodent pilot with multichannel recording at the stimulation site, designed to answer the same question. Experiment 3 is the closed-loop test: policies conditioned on nothing (fixed-rate), phase only, latent only, and phase-plus-latent, evaluated in the gym — including a non-stationary variant with drifting coupling, where the static-lag phase controller should degrade and the latent-conditioned controller should adapt — using the suppression coefficient (the strongest reported number being $S \approx 33$ with convergence within about 1,100 steps Reinforcement Learning Framework for Deep Brain Stimulation Study) plus a coherence-disruption metric and a pulse-count budget. The decisive check across all three experiments is the same falsifiable statement: if phase alone accounts for nearly all predictable outcome variance, then the PRC is not an average fiction and latent-state conditioning is unnecessary — and every experiment above is built to detect that outcome as cleanly as to detect its opposite.
References
- The Phase of Ongoing EEG Oscillations Predicts Visual Perception
- The Phase of Ongoing Oscillations Mediates the Causal Relation between Brain Excitation and Visual Perception
- Non-invasive suppression of essential tremor via phase-locked disruption of its temporal coherence
- Single reach plans in dorsal premotor cortex during a two-target task
- Discrete Stepping and Nonlinear Ramping Dynamics Underlie Spiking Responses of LIP Neurons during Decision-Making
- Amplification of Trial-to-Trial Response Variability by Neurons in Visual Cortex
- State-dependence of sensory-evoked responses in neocortex
- An orderly single-trial organization of population dynamics in premotor cortex predicts behavioral variability
- Response of Neuronal Populations to Phase-Locked Stimulation: Model-Based Predictions and Validation
- Intelligent Neural Interfaces: An Emerging Era in Neurotechnology
- Inferring single-trial neural population dynamics using sequential auto-encoders
- Reinforcement Learning Framework for Deep Brain Stimulation Study
- Neural Latents Benchmark '21: Evaluating latent variable models of neural population activity