When we record neural activity to understand how the brain thinks, decides, or remembers, we face a fundamental challenge: the brain is housed inside a physical body that is constantly moving. Whether an animal is actively running, subtly shifting its posture, or fidgeting during a delay period, these physical actions generate massive waves of neural activity.
Historically, neuroscientists often assumed that cognitive processes—like calculating the value of a choice or maintaining a task rule—could be isolated by averaging neural activity across trials. However, a growing body of research demonstrates that spontaneous, uninstructed, or preparatory movements heavily contaminate neural signals across the entire brain. This phenomenon, known as behavioral confounding, means that what looks like a neural representation of an abstract thought is often just the brain controlling a muscle or responding to physical arousal.
To map the functional circuits of the mind, researchers must find ways to decouple internal cognitive calculations from external motor execution.
The Ubiquity of Motor Contamination
Why is behavioral confounding so pervasive? The brain is an interconnected, embodied control system. Even in areas traditionally dedicated to pure sensory processing or high-level planning, motor signals are deeply intertwined with cognitive variables.
For instance, consider the primary visual cortex (V1). While V1 is classically viewed as a sensory relay for visual inputs, Márton A. Hajnal et al. found that its neural population activity is heavily influenced by physical movements and changes in internal arousal. This correlation makes it incredibly difficult to isolate true internal task representations from the neural variance caused by physical behavior. Furthermore, John P Liska et al. demonstrated that behavioral states like locomotion modulate the visual cortex in fundamentally different ways across species, highlighting how tightly physical activity and internal alertness are correlated.
Moving deeper into cognitive zones, the same problem appears. When animals perform memory-guided tasks, they often engage in uninstructed micro-movements or preparatory fidgeting. Susu Chen et al. showed that this behavioral noise heavily influences large-scale neural recordings in cognitive areas, making it unclear whether distributed delay-period activity represents an internal cognitive plan or merely the physical anticipation of an action. Similarly, Charles Findling et al. discovered that internal "prior expectations" are represented across nearly the entire mouse brain, but these signals are deeply entangled with embodied signals like preparatory motor movements and posture changes.
Even within the motor cortex itself, signals are confounded. The motor cortex does not just execute movements; it also processes cognitive variables like reward and target location. However, Sam Cooler argues that identifying reward-related neural activity in motor regions is severely complicated by kinematic confounding, where physical reactions to success or failure (such as frustration movements) mimic pure evaluative signals.
[Cognitive Variable] (e.g., Task Context, Value) \
--> [Shared Neural Population] --> Confounded Recording
[Behavioral Variable] (e.g., Running, Fidgeting) /
Experimental Decoupling via Temporal Separation
One of the most intuitive ways to untangle decision-making from motor execution is to separate them in time. By introducing forced delays between a sensory stimulus and the required motor response, researchers can allow sensory-cognitive signals to settle before the storm of motor activity begins.
Justin D. Yao et al. utilized this strategy to study how the parietal cortex transforms acoustic information into decision variables. Because late-stage neural activity often overlaps with physical motor preparation, incorporating forced delays allowed the authors to isolate pure sensory integration from the downstream motor commands. In a similar vein, Vahid Esmaeili et al. tracked the cortical choreography of a delayed sensory-triggered response in mice. By enforcing a waiting period after a whisker touch, they could separate active suppression signals—essential for motor memory—from the neural activity generated by preparatory fidgeting.
Geometric Disentanglement in Neural Subspaces
While temporal delays work well for structured tasks, natural behavior is fluid. To isolate cognitive signals during ongoing behavior, researchers leverage the concept of neural subspaces. If the brain records different types of information in independent, orthogonal dimensions of population activity, we can mathematically project the neural data to isolate the cognitive signal from the motor noise.
Neural Dimension Y (Motor Activity / Running)
^
| . (Running + High Context)
| . (Running + Low Context)
|
|---------------. (Still + High Context)----> Neural Dimension X
| . (Still + Low Context) (Pure Cognitive Context)
+---------------------------------------->
The brain itself appears to use this geometric strategy to prevent internal confusion. Márton A. Hajnal et al. discovered that the primary visual cortex multiplexes task rules and sensory data into orthogonal neural subspaces. Because these subspaces are independent, the representation of task context can predict behavioral success without interfering with visual processing or being overwritten by movement-related variance.
Engineers and neuroscientists use a similar mathematical intuition to analyze data from freely moving subjects. For example, Neda Shahidi et al. combined wireless electrophysiological recordings with automated tracking in foraging macaques. By identifying a specific, low-dimensional subspace within the prefrontal cortex, they successfully decorrelated movement noise from the neural data, isolating the representations of internal strategic variables (like reward history and timing) during naturalistic behavior.
Computational De-noising and Latent Factor Models
When experimental design and simple linear projections are not enough, advanced statistical models can be trained to decompose neural variance into behaviorally relevant and irrelevant components.
One major challenge is that a substantial portion of recorded neural activity exists as a "null space" of background noise that does not contribute to the behavior under study. To address this, Moein Khajehnejad et al. introduced an end-to-end neural network framework that jointly models neural spike trains and behavioral data as inhomogeneous Poisson processes, effectively linking microscopic neural activity to macroscopic actions while filtering out behaviorally irrelevant background noise.
Similarly, unsupervised methods can accidentally capture dominant patterns of neural activity that are completely unrelated to the cognitive trait of interest. To isolate the specific signals driving behavior, Neil Gallagher et al. developed Cross-Spectral Factor Analysis (CSFA), a machine learning framework that decomposes multi-site brain signals into interpretable factors, isolating neural synchronization patterns specifically associated with external variables rather than general signal variance.
For large-scale imaging data, Sayed Mohammad Hosseini introduced SBIND, a deep learning framework designed to disentangle behaviorally relevant brain activity from internal maintenance or unrelated background processing. By separating these dynamics, such models significantly improve our ability to decode intent and predict future neural states.
Finally, behavioral confounding can occur because we fail to categorize the behavior itself precisely. The authors of Distinguishing discrete and continuous behavioral variability using warped autoregressive HMMs developed WARHMM to solve a critical statistical conflation: the mixing of action selection (what an animal chooses to do) with action execution (how fast or vigorously it does it). By mathematically separating these two components using a latent warping variable, the model allows neuroscientists to determine whether a neural circuit regulates the cognitive choice of a behavior or its physical performance.
Why Disentanglement Matters
Resolving behavioral confounding is not just an academic exercise; it is a prerequisite for building reliable neural technologies and understanding neurological diseases.
In brain-machine interfaces (BMIs), positional information and decoding errors can act as major nuisance variables. Jonathan C. Kao et al. demonstrated that effector position can weakly affect motor cortical activity even in the absence of arm movements, presenting a confound for velocity decoders. If a BMI cannot distinguish between a user's pure intent and the physical state of the body, the control loop breaks down.
By understanding how the brain naturally untangles these signals—whether through temporal separation, orthogonal subspaces, or latent factor segregation—neuroscientists can build decoders and computational models that truly read the mind, rather than just tracking the movements of the body.
Key Foundational Work
Specificity of Human Cortical Areas for Reaches and Saccades (2007) — contributed 1 concept elements to this direction.
Cortical Representation of Ipsilateral Arm Movements in Monkey and Man (2009) — contributed 1 concept elements to this direction.