🧠 neuroscience

Tracking the Fidelity of Internal Neural Representations with Error-In-Variables Regression

This paper introduces a nonlinear error-in-variables regression framework that quantifies the fidelity of internal neural representations by modeling neural activity as a function of latent variables that deviate from measured sensory and behavioral data, successfully recovering hidden dynamics and condition-dependent representational changes in both synthetic and real neural recordings.

Original authors: Garon, I., Keeley, S., Williams, A. H.

Published 2026-08-03

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Garon, I., Keeley, S., Williams, A. H.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain is a super-advanced GPS, constantly calculating where you are, which way you're facing, and how fast you're moving. To do this, it relies on a team of tiny, hardworking neurons that act like sensors, firing in patterns that correspond to the world outside. But here's the catch: the "map" inside your brain doesn't always match the "map" the rest of the world sees. Sometimes, the GPS drifts because of a weak signal, sometimes it gets confused by a dark room, and sometimes it just decides to take a shortcut based on a hunch. For a long time, scientists have tried to figure out exactly what these neurons are thinking by looking at the outside world (like video tracking a mouse's position) and assuming the brain's internal map is a perfect copy of that video. But what if the brain's map is actually a bit wobbly, or even completely different from what the camera sees? If we assume the camera is perfect, we might think the brain is broken or confused, when really, the brain is just doing its own thing.

This is where a new study steps in to fix the misunderstanding. The researchers, working in the field of neuroscience, wanted to build a better way to listen to the brain's internal thoughts without getting tricked by the noisy, imperfect data we can measure from the outside. They developed a clever statistical tool that acts like a detective, separating the "true" internal feeling of a neuron from the "noisy" external measurement. By doing this, they can see if a neuron is actually losing its tune or if it's just that our measuring tape is slipping. This matters because understanding how the brain's internal map drifts away from reality helps us understand how we navigate the world, how we remember things, and what happens when our internal compass starts to fail.

The Detective's New Tool: Error-in-Variables Regression

The paper introduces a new method called Error-in-Variables (EIV) regression. Think of it like trying to guess a singer's true voice while standing in a room with a terrible echo and a shaky microphone. If you just record the sound and try to figure out the singer's pitch, you might think they are singing off-key because of the echo. But what if you could mathematically "subtract" the echo and the microphone wobble to hear the singer's actual voice? That is exactly what this new tool does for brain cells.

In the past, scientists used "conventional tuning curves." Imagine you are trying to draw a map of a city based on a blurry, shaky photo. If the photo is clear, your map looks great. But if the photo is blurry (maybe because the camera is moving or the lighting is bad), your map gets smeared out, and you might think the streets are wider or the buildings are in the wrong place. The paper shows that when scientists used these old methods on brain data, they often saw "smeared" maps and concluded that the neurons were getting confused or losing their precision.

The new EIV method says, "Wait a minute! Maybe the neurons are actually sharp and precise, but the measurement is the one that's blurry." It introduces a hidden variable—a "latent" internal state—which represents what the brain actually thinks is happening. The model then figures out how much the real world (the video camera) matches this internal thought. It uses a special dial called $\kappa$ (kappa) to control this match.

If $\kappa$ is high, the camera and the brain are perfectly in sync. The model acts like a standard, supervised map.
If $\kappa$ is low, the camera and the brain are totally out of sync. The model acts like an unsupervised detective, ignoring the camera and just looking at the brain's patterns to find its own map.
The magic is that the model can automatically tune this dial using a technique called cross-validation. It looks at the data and asks, "What setting of $\kappa$ makes the most sense?"

Testing the Theory with Fake Brains

Before trusting the tool with real animals, the authors tested it on synthetic data—a computer simulation of a brain. They created a fake scenario with 30 neurons and 1,000 time steps. They set up three different levels of "fidelity" (how well the internal map matched the external world):

High Fidelity ( $\kappa = 8.0$ ): The internal map and the camera matched almost perfectly.
Medium Fidelity ( $\kappa = 3.0$ ): There was some drift and noise.
Low Fidelity ( $\kappa = 0.5$ ): The internal map was very different from the camera.

In these simulations, the "true" tuning of the neurons (how they responded to the internal map) was kept exactly the same. But as the noise in the camera increased, the old "conventional" maps got flatter and flatter, looking like the neurons had lost their ability to tune in. However, the new EIV model successfully recovered the sharp, true tuning curves in all three cases. It correctly identified that the neurons hadn't changed; only the relationship between the brain and the camera had. The model also accurately guessed the value of $\kappa$ that was used to create the fake data, proving it could find the right level of supervision on its own.

Real-World Test 1: The Mouse's Head Direction

The researchers then applied this tool to real data from the anterodorsal thalamic nucleus (ADn) in mice, a brain region famous for acting like a compass. They recorded mice in three different situations:

Light: The mouse could see everything.
Dark: The mouse was in the dark, relying on memory and other senses.
Head-Fixed: The mouse was strapped down on a floating ball, so it couldn't move its head, but the room spun around it. This messed up its sense of balance (vestibular input).

In the light, the mouse's internal compass and the camera matched up well. The model found a high $\kappa$ , meaning the brain was tracking the world accurately.
In the dark, the match got worse. The model found a lower $\kappa$ , showing that the mouse's internal sense of direction was starting to drift away from where the camera saw it.
In the head-fixed condition, the drift was huge. The model found a very low $\kappa$ . The mouse's internal compass was spinning at a different speed than the room, or getting confused, even though the neurons themselves were still firing in a very organized, ring-like pattern.

Here is the big discovery: When scientists used the old methods, they saw the tuning curves flatten out in the dark and head-fixed conditions and thought the neurons were getting "noisy" or "broken." But the EIV model showed that the neurons were actually still sharp and precise. They were still tuned to the mouse's internal sense of direction. The problem wasn't the neurons; it was that the mouse's internal sense of direction had decoupled from the real world. The brain was still doing its job, but the "map" it was using was no longer aligned with the "camera."

The study also looked at how this mismatch changed over time. They found that in the head-fixed condition, the error (the difference between the internal map and the camera) didn't just stay random; it started to oscillate like a wave. Sometimes the brain and the world were close, and sometimes they drifted far apart, creating a rhythmic pattern of getting lost and finding the way again.

Real-World Test 2: The Rat's Grid Cells

Next, they tested the tool on grid cells in rats, which are neurons that fire in a honeycomb pattern as the animal moves through space. They recorded rats in a circular arena, first in the dark and then in the light with a visual cue.

Again, the old methods showed that the rats' position maps looked blurry and less informative in the dark. But the EIV model revealed that the internal representation of the rat's position was actually much sharper than the blurry maps suggested.

Spatial Information: The model calculated that the neurons carried more information about the rat's internal position than about the measured position, especially in the dark.
Clustering: When they looked at where the spikes happened, the old method showed the spikes were scattered in the dark. The EIV method showed that relative to the rat's internal map, the spikes were tightly clustered, just like in the light.

The model also compared three types of approaches:

Fully Supervised (ignoring the mismatch): This forced the brain's map to match the camera, resulting in a blurry, scattered map.
Fully Unsupervised (ignoring the camera completely): This found a clear map, but it was rotated or shifted so much that it didn't look like the rat's actual movement at all.
EIV (The Sweet Spot): This found the perfect middle ground. It kept the map sharp and clustered but kept it aligned enough with the rat's actual movement to be useful.

Decoding the Future

Finally, the team asked: "If we use this new tool, can we predict where the rat is better than before?" They tried to decode the rat's position from the brain activity alone, without looking at the camera.

In the light, the new method was only slightly better than the old one because the camera was already a good guide.
In the dark, the new method was significantly better. By tuning the $\kappa$ dial to the right spot, the model could predict the rat's position more accurately than the old methods, which were confused by the noise.

They also checked if the model was consistent. If they split the neurons into two groups, did both groups agree on where the rat was? Yes. The EIV model showed that even in the dark, the two groups of neurons were telling the same story about the rat's internal position, whereas the old methods suggested they were confused.

The Takeaway

This paper doesn't just offer a new math trick; it offers a new way of thinking. It suggests that when our measurements of the brain look messy or "smeared," it might not be because the brain is broken. It might be because the brain's internal map has drifted away from the external world, and we were trying to force them to match. By using this Error-in-Variables regression, scientists can finally see the brain's true, sharp signals even when the world around it is noisy, dark, or confusing. It turns out the brain is often more precise than we give it credit for; it's just that sometimes, it's navigating a different map than the one we are holding.

Technical Summary: Tracking the Fidelity of Internal Neural Representations with Error-In-Variables Regression

The Problem

Neuroscientists frequently seek to identify sensory and behavioral variables that correlate with neural activity. However, these correlations are often imperfect. Internal neural representations can systematically deviate from externally measured variables due to several factors:

Measurement Noise: Experimental measurements (e.g., video tracking of body pose) often contain errors.
Path Integration Errors: Navigational circuits may accumulate errors over time, causing internal estimates of position or heading to drift from reality.
Systematic Biases: Neural populations may represent external variables with offsets (e.g., representing past or future locations) or spontaneously switch reference frames.
Contextual Variability: Identical stimuli may elicit different responses due to subtle, unmeasured variations in trial conditions.

Current methodologies lack a principled statistical framework to quantify these mismatches. Supervised encoding models (e.g., Generalized Linear Models) typically treat external variables as noise-free inputs, leading to misleading conclusions when this assumption is violated. Conversely, unsupervised manifold inference models can extract latent variables but often struggle with interpretability, tuning, and data-limited regimes. There is a need for a method that bridges these extremes to rigorously characterize the "fidelity" of internal representations relative to external measurements.

Methodology: Error-In-Variables (EIV) Regression

The authors introduce a nonlinear Error-In-Variables (EIV) regression framework. This approach explicitly models neural activity as a function of latent internal variables ( $x_t$ ) that deviate from measured sensory/behavioral variables ( $s_t$ ).

Core Model Structure:

Latent Variable: The true internal representation is denoted as $x_t$ .
Observed Variable: The measured behavior/stimulus is $s_t$ .
Error Term: The relationship is modeled as $s_t = x_t + \epsilon_t$ , where $\epsilon_t$ represents the mismatch or noise.
Neural Likelihood: Neural activity ( $Y_{nt}$ ) is modeled as a function of the latent variable: $Y_{nt} \approx f_n(x_t)$ .
Fidelity Parameter ( $\kappa$ ): The conditional distribution of $s_t$ $s_{t}$ given $x_t$ $x_{t}$ is modeled (e.g., using a von Mises distribution for circular variables) with a concentration parameter $\kappa$ $κ$ .
- High $\kappa$ : $s_t \approx x_t$ (High fidelity; the model behaves like a supervised encoding model).
- Low $\kappa$ : $s_t$ is uninformative about $x_t$ (Low fidelity; the model behaves like an unsupervised manifold learner).
- Intermediate $\kappa$ : The model interpolates between supervised and unsupervised regimes.

Inference and Implementation:

Flexible Basis Expansion: The tuning functions $f_n(\cdot)$ are modeled using a weight-space formulation of Gaussian Processes (GP), employing a truncated Fourier basis set to approximate nonlinearities.
Sampling-Based Inference: Due to the intractability of the marginal likelihood integral, the authors use sampling-based approximations. They employ Quasi-Monte Carlo (Roberts sequence) for independent time points and Sequential Monte Carlo (Particle Filters) to incorporate temporal smoothness in the latent trajectory.
Hyperparameter Selection: The fidelity parameter $\kappa$ is tuned via cross-validated marginal likelihood, allowing the model to automatically determine the optimal degree of supervision for a given dataset.

Key Contributions

Unified Framework: The paper provides a single probabilistic framework that continuously interpolates between fully supervised tuning curve analysis and fully unsupervised manifold inference (specifically, Gaussian Process Latent Variable Models).
Quantification of Fidelity: It introduces a principled method to estimate the representational fidelity parameter $\kappa$ , quantifying how closely internal representations track external measurements.
Moment-by-Moment Residuals: The model yields time-resolved estimates of the mismatch ( $\epsilon_t = x_t - s_t$ ), enabling the analysis of how representational errors evolve dynamically.
Robustness to Noise: By decoupling the latent manifold from noisy observations, the method recovers sharper tuning curves and more accurate latent trajectories than conventional methods when measurement noise or representational drift is present.

Results

The framework was validated on synthetic data and applied to two distinct experimental datasets:

1. Synthetic Data Validation:

The model accurately recovered ground-truth latent dynamics, tuning curves, and the true fidelity regime ( $\kappa$ ) across varying levels of noise.
Cross-validation successfully identified the generative $\kappa$ value, demonstrating the data-driven capability to tune supervision.

2. Mouse Anterodorsal Thalamic Nucleus (ADn) - Head Direction:

Experimental Conditions: Mice were recorded in three conditions: well-lit (full cues), dark (no visual cues), and head-fixed (visual cues only, no vestibular input).
Fidelity Changes: Cross-validated $\kappa$ decreased progressively from light to dark to head-fixed conditions, indicating increasing mismatch between internal head direction and measured behavior.
Tuning Stability: While conventional tuning curves flattened significantly in dark and head-fixed conditions, the EIV-derived tuning curves to the latent variable remained stable. This suggests the neural manifold itself is preserved, but its alignment with external measurements degrades.
Dynamics: Analysis of residuals revealed distinct dynamical regimes. In the head-fixed condition, residuals exhibited autonomous oscillatory behavior, suggesting a decoupling of the internal representation from environmental signals.

3. Rat Medial Entorhinal Cortex (MEC) - Grid Cells:

Experimental Conditions: Rats navigated a circular arena in dark and light conditions.
Improved Spatial Information: EIV models recovered sharpened spatial tuning curves compared to conventional rate maps, which were blurred in the dark.
Spatial Clustering: Metrics like Ripley's L showed that spatial clustering of spikes was preserved when evaluated against the EIV latent variable, even when it appeared degraded against observed position.
Decoding Performance: EIV models with optimally tuned $\kappa$ outperformed both fully supervised (conventional Bayesian) and fully unsupervised decoders in predicting position on held-out data, particularly in the dark condition.
Theta Rhythm: The residuals between latent and observed position in the EIV model exhibited cleaner theta-frequency modulation (6–12 Hz) compared to unsupervised models, supporting the biological plausibility of the inferred latent space.

Significance and Claims

The authors claim that EIV regression is a powerful and computationally tractable tool for systems neuroscience. Its primary significance lies in its ability to:

Disentangle Noise from Signal: Distinguish between a degradation of the neural manifold itself versus a degradation in the alignment between the manifold and external measurements.
Provide Interpretability: Unlike fully unsupervised models where latent variables may drift from behavioral correlates, EIV models maintain a principled, inferable connection to experimental variables.
Enable New Analyses: The framework allows for the study of condition-dependent changes in representational fidelity, tuning gain, and profile, as well as the uncovering of latent population manifolds that are obscured in conventional analyses.

The paper concludes that this approach offers a rigorous method for tracking the fidelity of internal neural representations, providing a "sweet spot" between the rigidity of supervised models and the ambiguity of unsupervised ones. It suggests that many observed changes in tuning (e.g., flattening in low-light conditions) may reflect a decoupling of internal estimates from external reality rather than a loss of neural coding precision.