🤖 machine learning

Torus Graphs for Large Scale Neural Phase Analysis

This paper introduces a scalable stochastic score matching procedure for Torus Graph models that reduces computational complexity from $\mathcal{O}(d^{6})$ to $\mathcal{O}(d^{2})$ , enabling the inference of large-scale, state-dependent, and directional phase relationships in neural signals through novel Hidden Markov and autoregressive extensions.

Original authors: Jack Goffinet, Casey Hanks, David E. Carlson

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jack Goffinet, Casey Hanks, David E. Carlson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your brain is a massive orchestra with thousands of musicians (neurons) playing different instruments. To understand how they communicate, scientists often look at the timing of their notes, known as "phase." Just like a conductor keeps everyone in sync, these brain rhythms coordinate thoughts, sleep, and movement.

For a long time, trying to map who is listening to whom in this orchestra was like trying to understand a symphony by only listening to two instruments at a time. If the violin and the cello sound similar, we might think they are talking to each other. But what if they are both just copying the conductor? Traditional tools couldn't tell the difference between a direct conversation and a shared influence, and they could only handle a tiny handful of musicians (about 100) before getting overwhelmed.

This paper introduces a new, super-powered toolkit called Torus Graphs (TG) that finally lets us listen to the entire orchestra at once, even when it has nearly 2,000 musicians playing at different speeds.

Here is a breakdown of their three main innovations:

1. The "Stochastic Score Matching" Shortcut

The Problem: The original math for mapping these brain rhythms was incredibly heavy. It was like trying to solve a giant jigsaw puzzle by calculating the exact shape of every single piece's edge before you could even try to fit them together. This took so much computer memory that it crashed when the puzzle got bigger than 100 pieces.

The Solution: The authors invented a "stochastic" (random sampling) method. Instead of calculating every single edge at once, they take small, random samples of the puzzle pieces to figure out the shape.

The Analogy: Imagine trying to guess the weather in a whole country. The old way was to measure the temperature at every single house (impossible). The new way is to check a few random houses in different towns and use that to build a reliable map.
The Result: This made the math 1,000 times faster and allowed them to analyze 1,860 different brain signals simultaneously without the computer crashing.

2. The "State-Changing" Detective (TG-HMM)

The Problem: The brain isn't static; it changes its "mode" constantly. When you are awake, the orchestra plays one style. When you are in deep sleep, it switches to a completely different style. The old tools treated the whole recording as one big, messy mix, missing these distinct shifts.

The Solution: They added a "Hidden Markov Model" (HMM) to the toolkit. This acts like a detective who can tell when the orchestra has switched conductors.

The Analogy: Think of a radio that plays different genres. One minute it's jazz, the next it's classical. A standard recorder just hears "music." This new tool can say, "Ah, right now we are in the 'Jazz' state, and the musicians are interacting differently than they were during the 'Classical' state."
The Result: They applied this to mouse sleep data. They found that during sleep spindles (brief bursts of activity during sleep), the brain doesn't just get "noisier." Instead, specific connections between the cortex and thalamus turn on and off in a very precise, sparse pattern that previous tools missed because they were looking at the "average" noise.

3. The "Directional Arrow" Finder (AR-TG)

The Problem: Knowing two musicians are playing in sync is easy. Knowing who is leading and who is following is hard. Did the violinist start the melody, or did the cellist? Traditional tools often just draw a line between them, implying a two-way street.

The Solution: They created an "Autoregressive" version of their model. This looks at the past to predict the future.

The Analogy: If you see a person drop a ball, and then you see another person catch it, you know the direction of the action. This model looks at the history of the brain waves to see if Signal A's past helps predict Signal B's future better than Signal B's own past.
The Result: They mapped out "information highways" in the brain. They found that during wakefulness, high-frequency signals flow one way, but during sleep, low-frequency signals flow differently. They identified specific one-way routes, like a signal going from the "prelimbic cortex" to the "striatum," which suggests a specific direction of influence that wasn't clear before.

The Big Picture

The authors tested these tools on fake data first to prove they worked, and then on real recordings from a mouse's brain.

Wakefulness vs. Sleep: They confirmed that high-frequency connections are stronger when awake, while low-frequency connections dominate during sleep (matching what we already knew).
The New Discovery: However, the new tools revealed hidden details. They showed that the "sleep state" isn't just a blur of noise; it has specific, sparse, and directional patterns of communication that were previously invisible.

In short, this paper provides a scalable, mathematical "super-lens" that allows scientists to finally see the complex, shifting, and directional conversations happening inside the brain's massive network, moving beyond simple "who is loud" to "who is talking to whom, and in what direction."

Technical Summary: Torus Graphs for Large Scale Neural Phase Analysis

Problem Statement
Oscillatory neural signals, such as EEG and local field potentials (LFPs), exhibit phase relationships that coordinate communication across brain regions. Modern recording technologies now capture high-dimensional data (hundreds of channels across many frequency bins), yet standard phase analysis methods remain restricted. Classical circular statistics and pairwise metrics like Phase Locking Value (PLV) fail to scale beyond a few variables and cannot distinguish direct from indirect interactions, leading to spurious connectivity graphs. While the Torus Graph (TG) model was introduced as a principled exponential-family distribution capable of inferring conditional independence in multivariate phase data, its application has been limited to approximately 100 variables. This limitation stems from the $O(d^6)$ computational complexity of its standard score matching inference, which requires solving a linear system involving a $2d^2 \times 2d^2$ matrix.

Methodology
The authors propose a suite of scalable extensions to the Torus Graph framework to enable analysis of thousands of phase variables:

Stochastic Score Matching: To overcome the $O(d^6)$ bottleneck, the authors reformulate the score matching objective. By utilizing vector-Jacobian products (VJP) via reverse-mode automatic differentiation, they compute the necessary terms in $O(d^2)$ time per iteration without explicitly forming the dense Jacobian matrix. This reduces memory requirements from $O(d^4)$ to $O(d^2)$ and enables the use of stochastic optimization (Adam) with mini-batches.
Torus Graph Hidden Markov Model (TG-HMM): To capture state-dependent changes in phase coupling (e.g., sleep spindles), the authors introduce a TG-HMM. A central challenge in fitting HMMs with intractable normalizing constants (like TGs) is the infeasibility of standard Expectation-Maximization (EM) E-steps. The authors propose a "discriminative M-step" trick: they approximate the log normalizing constants with free parameters and use a logistic regression classifier to estimate soft state assignments. This allows for tractable EM steps where the M-step fits a discriminative model to approximate the stationary conditions of the exact M-step, avoiding the need to compute intractable partition functions.
Autoregressive Torus Graphs (AR-TG) and Transfer Entropy: To infer directional interactions, the authors develop AR-TGs, a conditional TG variant where covariates are lagged observations. This model supports the estimation of Transfer Entropy (TE) in multivariate phase systems. For multivariate TE, they employ a flexible imputation strategy (using a Gaussian model) to estimate the history of one time series from others, allowing for the calculation of conditional entropies required for TE estimation without fitting $O(C^2)$ separate models.

Key Contributions

Scalability: The introduction of stochastic score matching reduces the per-iteration cost of TG inference from $O(d^6)$ to $O(d^2)$ , enabling inference on datasets with thousands of variables (demonstrated up to $d=1,860$ ).
Dynamic Modeling: The TG-HMM provides a framework for capturing discrete, state-dependent changes in phase coupling structure, generalizing mixture models of torus graphs.
Directional Inference: The AR-TG framework enables the estimation of transfer entropy and directional dependencies in high-dimensional phase systems, moving beyond pairwise or static analyses.
Empirical Validation: The methods are validated on synthetic data for parameter recovery and causal discovery, and applied to real-world multi-electrode mouse LFP recordings.

Results

Synthetic Data: Stochastic score matching matches the accuracy of exact methods in low dimensions but significantly outperforms them in high dimensions where exact methods fail due to memory constraints. The TG-HMM accurately recovers hidden states, and AR-TGs demonstrate superior performance in identifying causal directions compared to baseline TE estimators and multivariate Granger causality, particularly in high-dimensional and sparse settings.
Neural Data (Mouse LFP): Applied to 48 hours of LFP data from 62 channels across 30 frequencies ( $d=1,860$ $d = 1, 860$ ):
- Large-Scale TGs: Revealed distinct phase interaction patterns between Wake and NREM sleep, showing stronger high-frequency couplings during Wake and low-frequency couplings during NREM. The learned parameters were substantially sparser than empirical statistics, reflecting conditional independence.
- TG-HMM: Applied to sleep spindles, the model identified a spindle-associated state characterized by sparse, spatially specific changes in conditional coupling (e.g., enhanced cortical interactions, reduced thalamic high-frequency coupling). This contrasted with PLV analysis, which showed diffuse, widespread synchrony increases, illustrating the value of removing indirect interactions.
- AR-TG: Revealed state-dependent directed motifs (e.g., prelimbic cortex $\to$ striatum) and frequency-specific routing (beta-band stronger in NREM, low-gamma stronger in Wake).

Significance and Claims
The paper claims to provide a general framework for the scalable, probabilistic modeling of circular-valued data. By enabling the analysis of phase relationships at previously inaccessible scales, these methods allow for the systematic mapping of dynamic and directional phase relationships across brain and cognitive states. The authors emphasize that while the frequency-specific reorganization observed between Wake and NREM validates the method against known physiology, the sparse conditional structures and directed asymmetries revealed are novel findings that suggest specific, testable hypotheses about state-dependent neural routing. The work is presented as a foundation for future extensions to other exponential families and probabilistic models, rather than a complete solution to all neural connectivity challenges. The authors note limitations, including the reliance on univariate targets for TE estimation and the fact that the inferred directed interactions reflect predictive (Granger-style) rather than interventional causality.

1. The "Stochastic Score Matching" Shortcut

2. The "State-Changing" Detective (TG-HMM)

3. The "Directional Arrow" Finder (AR-TG)

The Big Picture

Technical Summary: Torus Graphs for Large Scale Neural Phase Analysis

More like this