If you record the activity of a single neuron in the prefrontal cortex during a memory task, the data often looks like a chaotic mess. The neuron might fire vigorously at the start of a trial, fall silent, and then spike sporadically during the delay. For decades, neuroscientists treated this variability as "noise" to be averaged away across hundreds of trials.
However, a transformative perspective has emerged: this apparent mess is not noise, but a high-dimensional projection of a very simple, low-dimensional "engine" driving the circuit. By shifting focus from individual neurons to the collective activity of the population, researchers like Byron M. Yu and Krishna V. Shenoy have shown that neural activity is constrained to low-dimensional latent manifolds—smooth surfaces or trajectories that capture the essential computations of the brain.
The Manifold Hypothesis
The core intuition is that while a brain region may contain millions of neurons (the structural dimensionality), the number of independent ways those neurons can fire during a specific task (the functional dimensionality) is remarkably small.
In Extracting Dynamical Structure Embedded in Neural Activity, Byron M. Yu and colleagues demonstrated that motor planning in monkeys can be reconstructed on a single-trial basis by modeling neural activity as a low-dimensional latent trajectory. Instead of seeing 100 independent neurons, we see one "state" moving through a 3D space.
Why does this work? Eric M. Trautmann and Peiran Gao argue in A theory of multineuronal dimensionality, dynamics and measurement that the dimensionality of neural activity is fundamentally limited by the complexity of the task itself. If a monkey is only moving its arm in two dimensions, the neural "scaffold" supporting that movement doesn't need a thousand degrees of freedom.
Latent Variables as the Computational Engine
If neural activity lives on a manifold, what determines its movement? The field has moved toward treating the population as a latent dynamical system. In this view, the "state" of the brain at one millisecond predictably evolves into the state at the next, governed by internal rules.
Jakob H. Macke and John Patrick Cunningham showed in Empirical models of spiking in neural populations that these latent dynamical models are far more accurate than models that try to predict one neuron's firing based on its neighbors. This suggests that neurons aren't necessarily "talking" to each other in a simple web; rather, they are all tuned into a shared, hidden signal.
This principle extends beyond motor control:
- Decision Making: Mikhail Genkin and Tatiana A. Engel found that heterogeneous firing in the premotor cortex actually represents a simple, one-dimensional "decision variable" governed by attractor dynamics (The dynamics and geometry of choice in premotor cortex).
- Memory and Timing: Christian K. Machens and Ranulfo Romo demonstrated that the "messy" prefrontal cortex functionally separates "what" is being remembered from "when" it happened by projecting activity onto distinct, low-dimensional components (Functional, But Not Anatomical, Separation of “What” and “When” in Prefrontal Cortex).
Distilling Complexity into Tiny Models
If the underlying dynamics are truly low-dimensional, we should be able to emulate them with very small artificial systems. Recent work has proven this intuition correct. Li Ji-An and Marcelo G Mattar showed that Recurrent Neural Networks (RNNs) with as few as one to four units can outperform complex cognitive models in predicting human decision-making (Discovering cognitive strategies with tiny recurrent neural networks).
Similarly, Rylan Schaeffer and Ila Fiete reverse-engineered RNNs solving hierarchical tasks and found they compressed high-dimensional inputs into a 2D subspace where different axes represented different timescales of information (Reverse-engineering recurrent neural network solutions to a hierarchical inference task for mice).
The Limits of Compression
While the low-dimensional view is powerful, it faces a growing challenge. Are brain dynamics inherently low-dimensional, or is that an artifact of simple laboratory tasks?
In Motor cortex flexibly deploys a high-dimensional repertoire of subskills, researchers recorded from ~1,000 neurons and found that when tasks become more complex, the motor cortex reveals a much higher-dimensional repertoire of "subskills." This suggests that the brain's "manifold" might be more like a vast library of low-dimensional strategies that it can flexibly deploy, rather than a single, rigid constraint.
Understanding how the brain manages this balance—compressing information for efficiency while maintaining high-dimensional flexibility for complex behavior—remains the central frontier of the field.