Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining

Matthew J Bryan; Daniel C Muir; Felix Schwock; Azadeh Yazdan-Shahmorad; Rajesh P N Rao

cs.LGAug 27, 2026

Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining

Authors: Matthew J Bryan, Daniel C Muir, Felix Schwock, Azadeh Yazdan-Shahmorad, Rajesh P N Rao

Organizations: Neural Systems Laboratory, Paul G. Allen School of Computer Science & Engineering, University of Washington, Seattle, WA, USA · Center for Neurotechnology, University of Washington, Seattle, WA, USA · Computational Neuroscience Center, University of Washington, Seattle, WA, USA · Department of Electrical and Computer Engineering, University of Washington, Seattle, WA, USA ·

Abstract

Objective: Model-based closed-loop neural stimulation holds promise for therapeutic applications ranging from Parkinson's disease to sensory restoration, but deployment has been limited by two obstacles: 1) forecasting models for predicting the consequences of stimulation fail catastrophically on a meaningful fraction of sessions, and 2) per-session calibration requirements are often incompatible with clinical constraints. We address both by demonstrating, for the first time, that meta-learning and pretraining can be applied to neural stimulation response modeling. Methods: Temporal basis function models (TBFMs) forecast state-dependent neural responses to stimulation. We extend TBFMs with cross-session pretraining using a novel architecture and algorithm based on model-agnostic meta-learning (MAML), evaluating them on 40 sessions of optogenetic stimulation in primary sensorimotor cortex of two non-human primates. Results: Meta-learning substantially reduces catastrophic forecast failure: for a 1k calibration set size, sessions with test R-squared < 0.05 drop from 16 of 40 (single-session training) to 1 (MAML-pretrained), and prediction intervals become significantly narrower (p < 0.05). Calibration requirements are reduced by 50-90% at matched accuracy, enabling experiments otherwise infeasible within clinical session-time constraints. Conclusion: Our results demonstrate that cross-session structure in stimulation responses is consistent enough to support pretraining, providing the first empirical evidence that meta-learning approaches are viable for neural stimulation. Significance: The robustness and sample efficiency gains directly address known obstacles to deploying model-based stimulation controllers. Our results motivate community efforts to assemble standardized multi-site stimulation datasets and to further explore meta-learning for robust closed-loop stimulation.

Figures & tables

Fig. 1: Stimulation response across the $\mu$ ECog array. Heatmap indicates LFP values (Gaussian-smoothed and Z-scored, single session average) at 10ms after the second stim pulse onset. Missing electrodes were given false values for graphing. Stim locations varied between sessions.

Fig. 2: Data windowing Trials are 184ms separated by a 16ms buffer, $\sim 15k$ trials per session. See Figure 3 for legend.

Fig. 3: Average response across many trials for a single electrode within one session. The TBFM model predicts the stimulation-evoked activity ( $y$ ) in each trial using the first 20 ms of recorded neural data ( $x$ , known as the “runway”). The initial optical pulse occurs at 40 ms (blue dashed line); we assume there is a 20 ms control-loop delay between measuring the brain state at 20 ms and computing parameters for the stimulation (see Sec. II-B1 ). Note that the TBFM forecasts all channels simultaneously using the runway for all channels, but we depict only one channel here.

Fig. 5: Mean neural response and mean prediction for four binned initial states. Single channel, single session. Stimulation onset is at t= $40ms$ . $y$ is the mean stimulation response for trials in the given initial state bin, $\hat{y}$ is the mean forecast for the same trials.

Fig. 6: Multi-session TBFM architecture. $X_{b,t,d}$ Neural activity in a session (each block depicted is one of $n$ sessions), time span $t$ , channel count $d_{*}$ , multiple trials per session: $b$ is the batch dimension. $r$ is the runway length, $T$ is the forecast horizon. $Z_{t,l}$ Latent space representation of neural activity for a session, time span $t$ , latent dimensionality $l$ . $s_{i}$ per-trial embedding of stimulation parameters, known as the “stimulation descriptor”. $c^{rest}_{s}$ per-session context from resting/non-stimulation activity, repeated for all trials of that session. $c^{stim}_{s}$ learned per-session stimulation context, repeated for all trials of that session.

Fig. 7: Example deployment . In this simple example we collect larger calibration (randomly timed stimulation) data sets during the initial 20 sessions. During those sessions we create a single-session TBFM for closed-loop experimentation. After 20 sessions, we use a meta-learned shared-session TBFM and test-time adaptation, allowing us to collect far smaller calibration data sets (5k versus 500 calibration examples).

Fig. 8: Distribution of test set $R^{2}$ across folds and sessions, by calibration set size and training method. Dashed lines indicate median and IQR. Left tails of the distributions are significantly heavier for vanilla (i.e., single-session) models, indicating greater robustness of the coadaptive- and MAML-pretrained models. At the calibration set size of $1k$ for example, 7 of 40 sessions with coadapted model, and 7 of 40 with MAML models have test set $R^{2}<0.15$ , compared with 18 of 40 for vanilla models. At $2.5k$ it is 8, 6 and 11, respectively.

Fig. 9: A-ACF predicts performance of pretrained models. Robustness is indicated by high gains on noisy sessions. Lower A-ACF values on resting data suggest a session’s data are noisier. The median A-ACF value across a session’s channels is predictive of both vanilla (single-session) model and MAML TTA model performance: the trend slope is positive at $p<5^{-5}$ in all cases. Pretraining flattens the susceptibility of the model to failing catastrophically on high noise (low A-ACF) sessions, and hence we see large improvements in the high noise regime. Vanilla models more frequently fail catastrophically ( $R^{2}<0.05$ ) on higher noise sessions. At 5k calibration trials (right), the two methods converge, indicating that pretraining’s advantage is specific to low-data regimes where anchoring to cross-session patterns compensates for limited per-session evidence. The slope of the difference in trends is negative at $1^{-3}$ for 1k, insignificant in 5k. The 5k result is expected since prior work found this was the optimal training set size for single-session models [ 7 ] .

Fig. 10: Sample efficiency of single session versus MAML pretrained models. The plot shows mean and distribution of test set $R^{2}$ across held out sessions and Monte Carlo Cross-Validation (MCCV) splits. Confidence intervals (CIs) are for the mean of held-out sessions. The prediction intervals (PIs) capture the variability in performance across sessions and MCCV splits, reflecting both genuine cross-session differences in signal quality and the uncertainty introduced by the particular held-in/held-out assignment. Mean performances of MAML-pretrained models exceed vanilla (single session) models at $p<0.05$ under the 1-sided Wilcoxon signed-rank test for calibration set sizes $\leq 2.5k$ , and $p=0.15$ for $5k$ . Prediction intervals are significantly more narrow for MAML-pretrained models under the single-sided Brown-Forsythe test ( $p<0.05$ ) for calibration set sizes $\leq 2.5k$ , indicating more consistent performance and higher robustness. At $5k$ , it drops to $p=0.19$ .

	Held out subject
Held in subject	G	J
G	0.405 ± 0.052	0.354 ± 0.094
J	0.407 ± 0.047	0.373 ± 0.102

TABLE I: Summary of Within- and Cross-Subject Generalization. Values are held-out session test set performance, mean across folds. Indicated interval is the

95\%

confidence interval of the mean.

	CPU	GPU
Pretrained TBFM	2.798	0.790
Pretrained TBFM compiled	2.095	0.223
Vanilla TBFM	0.174	0.226
Vanilla TBFM compiled	0.115	0.121
AE-LSTM	41.628	34.156
LSSM	10.111	16.939

TABLE II: Average single-trial inference times (milliseconds)

Neural Surrogate Modeling

Neural surrogate modeling uses learned models to approximate costly simulations, enabling faster prediction, design, control, and inference.

See all 418 papers

Apr 13, 2026

Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching

Forecasting neural activity in response to naturalistic stimuli remains a key challenge for understanding brain dynamics and enabling downstream neurotechnological applications. Here, we introduce a generative forecasting framework for modeling neural dynamics based on autoregressive flow matching (AFM). Building on recent advances in transport-based generative modeling, our approach probabilistically predicts neural responses at scale from multimodal sensory input. Specifically, we learn the conditional distribution of future neural activity given past neural dynamics and concurrent sensory input, explicitly modeling neural activity as a temporally evolving process in which future states depend on recent neural history. We evaluate our framework on the Algonauts project 2025 challenge functional magnetic resonance imaging dataset using subject-specific models. AFM significantly outperforms both a non-autoregressive flow-matching baseline and the official challenge general linear model baseline in predicting short-term parcel-wise blood oxygenation level-dependent (BOLD) activity, demonstrating improved generalization and widespread cortical prediction performance. Ablation analyses show that access to past BOLD dynamics is a dominant driver of performance, while autoregressive factorization yields consistent, modest gains under short-horizon, context-rich conditions. Together, these findings position autoregressive flow-based generative modeling as an effective approach for short-term probabilistic forecasting of neural dynamics with promising applications in closed-loop neurotechnology.

Nicole Rogalla, Yuzhen Qin, Mario Senden +2

Department of Machine Learning and Neural Computing, Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, the Netherlands · Department of Cognitive Neuroscience, Maastricht University, Maastricht, the Netherlands · Maastricht Brain Imaging Center, Maastricht University, Maastricht, the Netherlands

Functional Magnetic Resonance Imaging Probabilistic Time Series Forecasting Conditional Generative Modeling Flow-Based Generative Modeling

Ask PDF

May 11, 2026

CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients

Motor imagery electroencephalography (MI-EEG) decoding offers a non-invasive route for post-stroke rehabilitation, but cross-patient use remains difficult because pathological neural reorganization changes task-related EEG dynamics, aperiodic activity, local excitability, cross-regional coordination, and trial-level brain-state context. This makes source-learned MI representations unreliable for unseen patients. To address this problem, we propose CFSPMNet, a cross-patient adaptation framework that models post-stroke MI-EEG as latent neural-state organization. CFSPMNet combines a Fourier-Reorganized State Mamba Network (FRSM) with Shared-Private Prototype Matching (SPPM). FRSM represents each trial as a latent physiological token sequence, reorganizes token states in the Fourier domain, and uses Fourier-derived trial context to guide Mamba state-space propagation. SPPM improves target pseudo-label updating by combining semantic confidence with shared-private physiological consistency, filtering confident but physiologically inconsistent target predictions. Leave-one-subject-out experiments on two stroke MI-EEG datasets show that CFSPMNet outperforms representative CNN-, Transformer-, Mamba-, and adaptation-based baselines, achieving average accuracies of 68.23% on XW-Stroke and 73.33% on 2019-Stroke, with gains of 5.63 and 8.25 percentage points over the strongest competitors. Ablation, sensitivity, feature-alignment, pseudo-label selection, and neurophysiological visualization analyses further support the roles of Fourier-domain token-state reorganization and calibrated pseudo-label updating. These results suggest that latent neural-state modeling can improve rehabilitation-oriented cross-patient BCI decoding. Code is available at https://github.com/wxk1224/CFSPMNet.

Xiangkai Wang, Yun Zhao, Dongyi He +5

School of Artificial Intelligence, Chongqing University of Technology, Chongqing 401135, China; Chongqing Key Laboratory of Embodied Intelligence Perception and Autonomous Learning for Humanoid Robots; Key Laboratory of Advanced Equipment Intelligence of the Chongqing Education Commission, Chongqing 401135, China · School of Smart Health, Chongqing Polytechnic University of Electronic Technology, Chongqing 401131, China · Department of Language Science and Technology, The Hong Kong Polytechnic University, Hung Hom, Hong Kong SAR, China +1

Cross-Subject EEG Decoding Brain-Computer Interfaces Domain Adaptation EEG Decoding Frequency-Domain Feature Learning

Ask PDF

Apr 3, 2026

CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data

Recent work suggests that large-scale, multi-animal modeling can significantly improve neural recording analysis. However, for functional calcium traces, existing approaches remain task-specific, limiting transfer across common neuroscience objectives. To address this challenge, we propose CalM, a self-supervised neural foundation model trained solely on neuronal calcium traces and adaptable to multiple downstream tasks, including forecasting and decoding. Our key contribution is a pretraining framework, composed of a high-performance tokenizer mapping single-neuron traces into a shared discrete vocabulary, and a dual-axis autoregressive transformer modeling dependencies along both the neural and the temporal axis. We evaluate CalM on a large-scale, multi-animal, multi-session dataset. On the neural population dynamics forecasting task, CalM achieves competitive performance against strong specialized baselines after pretraining. With a task-specific head, CalM further adapts to the behavior decoding task and achieves superior results compared with supervised decoding models. Moreover, linear analyses of CalM representations reveal interpretable functional structures beyond predictive accuracy. Taken together, we propose a novel and effective self-supervised pretraining paradigm for foundation models based on calcium traces, paving the way for scalable pretraining and broad applications in functional neural analysis. Code is released at https://github.com/TSuXinH/CalM.

Xinhong Xu, Yimeng Zhang, Qichen Qian +1

Department of Automation, Tsinghua University, Beijing, China · School of Life Sciences, Tsinghua University, Beijing, China

Self-Supervised Pre-Training Time-Series Foundation Models Neural Decoding

Ask PDF

Abstract

Neural Surrogate Modeling

Explore similar work

Probabilistic Prediction of Neural Dynamics via Autoregressive Flow Matching

CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients

CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data