SYNC26: Spatiotemporal dYnamics in Neural Computation

Goals

To explore how neural dynamics underlie cognition, memory, and perception. This interdisciplinary topic invites researchers from neuroscience, machine learning, applied math, physics and cognitive science to examine how spatiotemporal coordination in neural populations supports communication, gating, temporal binding, and distributed computation; and further, how analogous principles might be leveraged in AI. We hypothesize that combining computational modeling with real-data analysis will reveal distinct neural dynamics underlying normal as well as violation conditions, shedding light on how the brain transitions between expected and unexpected states.

Projects

1. Modeling and interpreting neural traveling waves: Develop models that reproduce cortical traveling waves and test how they support distributed computation and routing.

2. Wave-based state-space models for motor control: Apply wave SSMs to motor control tasks to investigate how oscillatory dynamics contribute to planning and adaptive feedback.

3. Adaptive thresholds in delta-sparse recurrent networks: Explore if adaptive thresholding improves efficiency.

4. Phase-coded representation and reasoning in AI: Investigate how phase-based coding could be introduced into large language models.

5. Marker identification on real data: Analyze electrophysiological recordings to identify wave and control markers.

6. Oscillation-based brain–computer interfaces: Design BCIs that decode intent, speech, or movement from oscillatory features (e.g., beta/gamma rhythms, ERPs), focusing on marker extraction and real-time closed-loop control.

Materials and Equipment

Talks and Tutorials

We will examine neural dynamics from complementary perspectives (perception, prediction, and control) using both empirical data and computational models. Our focus will be on understanding how the brain transitions between expected and unexpected states by integrating computational modeling with analyses of real electrophysiological data. Leveraging the expertise and advanced hardware available at Telluride, this work may include:

• System confusions and expectation violations: We aim to model perceptual violations, particularly in auditory tasks. Electrophysiological data, either newly recorded or drawn from existing datasets, will be analyzed to identify neural markers associated with these phenomena. These findings will inform computational simulations of oscillatory dynamics, enabling us to explore translation mechanisms that may extend to language model frameworks where applicable. Candidate paradigms include Shepard’s scale (ABR / circular manifolds), binaural beats (ASSR / phase locking), phoneme restoration illusion (ERP / Bayesian inference), and lexical order violations (ERP / RNN modeling). The task and model will be selected based on participant interest. Through this combined empirical and computational approach, we aim to uncover principles that connect perceptual anomalies with neural oscillations and predictive processing models.

• Peripersonal perception and anticipatory actions: Understanding how the brain generates anticipatory motor responses to approaching objects in peripersonal space is fundamental to both neuroscience and the development of adaptive robotic systems. We propose investigating how sensory predictions, instead of purely motor commands, drive anticipatory behaviours. We will model this process using a bi-directional hierarchical predictive coding framework, in which prediction and error signals continuously interact across layers representing sensory features, perceptual constructs, and bodily actions or outcomes. Because this recurrent system evolves toward equilibrium over time, its dynamics directly reflect how anticipatory states stabilize and how expectation violations trigger reorganization. Human experiments will measure behavioural and neural responses during interactions with approaching stimuli (e.g., auditory or visual) under both expected and unexpected conditions. These insights will then inform the implementation of predictive coding architectures, testing whether models can exhibit human-like anticipatory behaviours and adaptive responses to expectation violations.

• Control and modulation in recurrent networks: Recent advances in wave state-space models (SSMs) show that oscillatory latent representations can capture long-range dependencies more effectively than many current sequence models. Yet both biological and artificial systems rely on gating to selectively amplify or suppress signals to regulate memory and forgetting. We will explore ensembles of wave-based SSMs augmented with different gating mechanisms (coupling versus input) to test how these mechanisms influence performance on sequence processing benchmarks. Gating applied to ensemble coupling parameters will approximate neuromodulatory control in the brain, while input gating offers direct control over information flow. Comparing these regimes will clarify how gating shapes memory retention and enables adaptive transitions between dynamic states. In collaboration with the auditory group, we will further investigate travelling wave modelling of temporal coherence and explore paradigms of source segregation in different domains.

Recommended reading

Keywords: