When recording from large populations of neurons, neuroscientists face a fundamental challenge: individual neurons often exhibit mixed selectivity. A single neuron in the motor cortex might alter its firing rate based on the timing of a movement, the specific target direction, and the environmental context. When hundreds of such neurons are recorded simultaneously, they form a high-dimensional state space where these distinct behavioral and cognitive variables are deeply entangled.
Traditional unsupervised dimensionality reduction techniques, such as Principal Component Analysis (PCA), find a low-dimensional subspace by maximizing the total variance captured in the data. However, because PCA is blind to experimental design, its resulting axes typically lump multiple sources of variance together. This makes it difficult to isolate how the brain encodes specific, independent parameters. To solve this, researchers have developed demixed dimensionality reduction (and related targeted dimensionality reduction techniques), which explicitly incorporate task parameters to partition and isolate independent information streams within neural population activity.
The Intuition of Demixing
The core mechanism behind demixed dimensionality reduction is to inject knowledge of the experimental design directly into the mathematical decomposition. Instead of treating all trial-by-trial variance equally, the total neural activity is first broken down into marginalized components based on task variables.
Imagine an experiment where an animal performs a task with two varying parameters: stimulus identity and elapsed time. To isolate the pure effect of the stimulus, we can average the neural population activity across all time points for each specific stimulus. This process strips away temporal fluctuations, leaving a data matrix that reflects only stimulus-driven variance. Conversely, by averaging across all stimuli for each time point, we isolate the pure temporal dynamics of the population.
Once the data is partitioned into these task-specific components, the algorithm finds pairs of encoder and decoder matrices. Unlike standard PCA, which requires orthogonal projection axes, demixed methods allow the projection axes to be non-orthogonal. This flexibility is crucial: it allows the algorithm to find a specific axis that captures the variance of one task variable while completely ignoring (or remaining orthogonal to) the variance caused by other variables. As a result, high-dimensional population activity is mapped onto distinct, interpretable, and low-dimensional subspaces.
Disentangling Motor Intent from Context
The power of this approach is highly evident in the study of motor control, where movement execution is always embedded within a physical context.
For instance, [Kaufman] et al. applied demixed dimensionality reduction in The Largest Response Component in the Motor Cortex Reflects Movement Timing but Not Movement Type to challenge traditional tuning-centric views of the motor cortex. They discovered that the most dominant signal in the population does not encode the direction or shape of a movement at all; instead, it captures global temporal dynamics—the internal "trigger" telling the body when to move. By separating task-dependent components from task-independent components, [Borgognon] et al. similarly demonstrated in Regional specialization of movement encoding across the primate sensorimotor cortex that the premotor cortex focuses heavily on task-specific planning, whereas primary motor and somatosensory areas handle general execution and temporal sequencing.
Demixed methods also reveal how the brain maintains consistent motor commands across changing physical environments. In brain-computer interface (BCI) research, [Rastogi] et al. used demixed principal component analysis (dPCA) in The Neural Representation of Force across Grasp Types in Motor Cortex of Humans with Tetraplegia to isolate a core, decodable signal for force intent that remains invariant across different hand postures.
This modular organization appears repeatedly across motor areas:
- Vargas-Irwin (2025) used dPCA to show that the human precentral gyrus isolates the variance of an intended hand gesture from the specific limb (left vs. right arm) used to execute it.
- [Lin] et al. demonstrated in Representation and decoding of bilateral arm motor imagery using unilateral cerebral LFP signals that local field potentials can be demixed into independent dimensions representing movement laterality versus body region.
- [Mender] et al. showed in The Impact of Task Context on Predicting Finger Movements in a Brain-Machine Interface that demixed variance decomposition can separate context-dependent neural shifts (like muscle resistance) from core motor commands.
- To understand how these independent components combine, the authors of Pseudo-linear summation explains neural geometry of multi-finger movements in human premotor cortex projected population activity onto targeted axes to reveal that the brain sums individual finger signals through a non-linear normalization filter, preventing neural saturation within the shared manifold.
Tracking Internal Cognitive States
Beyond motor execution, demixed dimensionality reduction provides a window into purely internal cognitive operations, such as working memory, attention, and decision-making.
In cognitive tasks, multiple abstract variables must be processed simultaneously without interfering with one another. [Bae] et al. utilized demixed dimensionality reduction in Parallel processing of working memory and temporal information by distinct types of cortical projection neurons to prove that the prefrontal cortex delegates working memory maintenance and timing tracking to completely different projection neuron types. Similarly, [Lundqvist] et al. employed dPCA in Working memory control dynamics follow principles of spatial computing to demonstrate that abstract control signals (like item ordering) are encoded in neural spaces distinct from the identities of the items themselves. In the cerebellum, [Avila] et al. used dPCA in Purkinje Cell Activity in the Medial and Lateral Cerebellum During Suppression of Voluntary Eye Movements in Rhesus Macaques to isolate the cognitive planning required to suppress a movement from the motor execution itself.
By defining these precise, task-relevant axes, researchers can project single-trial neural activity onto them to track internal states in real time. For example, [Cohen] et al. showed in A Refined Neuronal Population Measure of Visual Attention that projecting population responses onto a targeted scalar axis captures moment-to-moment fluctuations in visual attention. [Snyder] et al. leveraged a similar projection approach in Distinct population codes for attention in the absence and presence of visual stimulation to isolate a predictive preparatory attention signal that operates independently of sensory processing.
This trial-by-trial isolation helps explain complex behavioral phenomena across diverse domains:
- Choking under pressure: [Smoulder] et al. used targeted axes in A neural basis of choking under pressure to show that exceptionally high rewards cause an exceptional collapse in the distinctiveness of motor preparation spaces.
- Foraging and internal models: [Shahidi] et al. used canonical correlation analysis in Population coding of strategic variables during foraging in freely moving macaques to isolate a low-dimensional subspace tracking reward history and temporal expectations.
- Visuomotor adaptation: [Kao] et al. demonstrated in Trial-by-Trial Motor Cortical Correlates of a Rapidly Adapting Visuomotor Internal Model that projecting single-trial activity onto an environmental parameter axis can predict upcoming reach errors before the movement begins.
- Sensory calibration: [Zaidel] et al. used targeted dimensionality reduction in Supervised Multisensory Calibration Signals Are Evident in VIP But Not MSTd to separate external sensory stimuli from internal decision variables, pinpointing exactly where multisensory learning alters neural representations.
- Representational drift: [Xia] et al. showed in Stable representation of a naturalistic movie emerges from episodic activity with gain variability that decomposing population activity into spatial, temporal, and gain factors explains how sensory representations remain stable over weeks because long-term fluctuations are confined to a non-coding subspace.
Ultimately, demixed dimensionality reduction works because it abandons the assumption that the most variable features in a dataset are the most biologically meaningful. By aligning the mathematical projections directly with experimental variables, it uncovers the structured, modular architecture the brain uses to compute, plan, and act.
Key Foundational Work
A Refined Neuronal Population Measure of Visual Attention (2015) — contributed 1 concept elements to this direction.
Demixing Population Activity in Higher Cortical Areas (2010) — contributed 1 concept elements to this direction.