Announcer The following program features simulated voices generated for educational and philosophical exploration.

Adam Ramirez Good evening. I'm Adam Ramirez.

Jennifer Brooks And I'm Jennifer Brooks. Welcome to Simulectics Radio.

Adam Ramirez Tonight we're exploring how the geometric structure of neural population activity constrains and enables computation. Single neurons provide limited information about brain function—understanding cognition requires analyzing how populations of neurons collectively represent and transform information. Recent work reveals that neural population activity occupies low-dimensional manifolds within the high-dimensional space defined by all recorded neurons. These manifolds have geometric structure—they're curved, twisted, folded—and this geometry appears to reflect computational operations. The manifold perspective suggests that understanding brain computation requires characterizing the geometric transformations that populations perform rather than tracking individual neuron responses. This has implications for interpreting neural recordings, designing brain-machine interfaces, and understanding how neural networks—biological and artificial—solve computational problems through collective dynamics.

Jennifer Brooks The manifold hypothesis proposes that despite recording from hundreds or thousands of neurons, population activity is constrained to much lower-dimensional spaces. Instead of neurons varying independently, their activity is coordinated such that the population state—the vector of all firing rates at a moment—lies on or near a curved surface embedded in the high-dimensional neural space. This dimensionality reduction could reflect various constraints: shared inputs, recurrent connectivity, or optimization for specific computations. The geometry of these manifolds—their dimensionality, curvature, and topology—potentially encodes task variables and constrains computational operations. Understanding this requires moving beyond traditional approaches that average across trials or analyze single neurons to methods that characterize population geometry and dynamics.

Adam Ramirez To examine how neural manifolds structure computation and what their geometric properties reveal about circuit function, we're joined by Dr. Krishna Shenoy, neural engineer at Stanford University, whose work characterized population dynamics during motor control and developed brain-machine interfaces exploiting manifold structure. Dr. Shenoy, welcome.

Dr. Krishna Shenoy Thank you. The manifold perspective fundamentally changes how we think about neural population coding and computation.

Jennifer Brooks What evidence demonstrates that neural populations occupy low-dimensional manifolds rather than exploring the full high-dimensional space?

Dr. Krishna Shenoy Dimensionality reduction analyses consistently show that population activity lies in spaces far lower than the number of recorded neurons. In motor cortex during reaching movements, hundreds of neurons have activity that can be captured by roughly ten dimensions. In visual cortex during image viewing, thousands of neurons span perhaps dozens of dimensions. Principal component analysis, factor analysis, and other methods reveal shared variance structure indicating coordinated activity. This isn't simply due to insufficient sampling—recording more neurons reveals similar intrinsic dimensionality. The manifold structure appears in spontaneous activity and during behavior. Moreover, when we perturb neural populations through stimulation or lesions, activity remains confined to the manifold rather than exploring arbitrary dimensions, suggesting the low-dimensional structure is maintained by circuit dynamics.

Adam Ramirez What causes this dimensionality reduction? Is it imposed by circuit architecture or emergent from optimization?

Dr. Krishna Shenoy Multiple factors contribute. Shared inputs create common fluctuations across neurons. Recurrent connectivity couples neurons so their activity becomes interdependent rather than independent. Noise correlations—shared trial-to-trial variability—reduce effective dimensionality. From a computational perspective, low-dimensional dynamics may be optimal for certain tasks. Motor control benefits from smooth, coordinated population trajectories rather than independent neuron fluctuations. Sensory processing may extract low-dimensional task-relevant features from high-dimensional inputs. The manifold structure likely reflects both architectural constraints—the connectivity patterns evolution and development created—and functional optimization for the computations those circuits perform. Distinguishing these requires comparing manifold structure across tasks and manipulating circuit properties.

Jennifer Brooks How does manifold geometry relate to represented variables and computational operations?

Dr. Krishna Shenoy The manifold geometry encodes information. In motor cortex, the manifold is curved and rotates during movement preparation and execution—different reach directions correspond to different trajectories along the manifold. The curvature and rotation reflect dynamical operations transforming preparation activity into movement commands. In visual cortex, object identity and position map to different manifold dimensions, with object transformations like rotation corresponding to movements along the manifold. The geometric structure—distances between points, local dimensionality, curvature—captures computational relationships. Linearly separable task conditions in behavior often correspond to geometrically separable manifold regions. Computational operations like categorization may correspond to projecting population states onto decision boundaries that respect manifold geometry. Understanding computation requires characterizing how manifolds transform across processing stages.

Adam Ramirez How stable are neural manifolds across time, across behavioral contexts, across learning?

Dr. Krishna Shenoy Manifold structure shows both stability and flexibility. Over timescales of days to weeks, the manifold geometry remains relatively stable even as individual neuron tuning drifts—the population preserves computational structure despite component instability. This stability enables robust decoding for brain-machine interfaces. However, manifolds adapt with learning. When animals learn new motor skills, the manifold dimensionality and geometry change—new dimensions emerge or existing dimensions reorganize. Across behavioral contexts, manifolds can rotate or shift while maintaining local structure—motor cortex manifolds differ between reaching and grasping but share geometric principles. This suggests a distinction between the manifold structure, which is relatively stable and reflects circuit architecture, and the population state trajectory along that manifold, which varies with behavior and learning.

Jennifer Brooks What methods are used to characterize manifold geometry beyond dimensionality reduction?

Dr. Krishna Shenoy Linear methods like PCA provide initial dimensionality estimates but assume manifolds are flat, which is often violated. Nonlinear methods like Isomap, Laplacian Eigenmaps, and t-SNE better capture curved manifolds by preserving local distance relationships. Dynamical systems approaches model how population states evolve through time, identifying fixed points, limit cycles, and slow subspaces. Topological data analysis characterizes manifold shape through persistent homology, revealing holes, loops, and higher-dimensional structures. Geometric methods compute curvature, tangent spaces, and geodesics on manifolds. Each method makes assumptions and has limitations. Combining approaches provides converging evidence about manifold structure. Validating that estimated manifolds reflect neural computation rather than analysis artifacts requires showing that manifold properties predict behavior, remain stable across sessions, and are disrupted by relevant perturbations.

Adam Ramirez How do neural manifolds in biological systems compare to those in artificial neural networks?

Dr. Krishna Shenoy Trained artificial networks also exhibit low-dimensional population dynamics and manifold structure. Networks solving similar tasks often develop similar manifold geometries regardless of specific connectivity or training, suggesting that task requirements shape manifold structure more than architectural details. This convergence provides evidence that manifolds reflect computational solutions to task demands. However, biological and artificial manifolds differ. Biological manifolds tend to be lower-dimensional relative to population size. Biological manifolds show more variability and noise. Artificial networks can be designed to have interpretable manifold structure, whereas biological manifolds emerge from complex, partially unknown connectivity. Comparing manifolds across systems helps distinguish universal computational principles from implementation-specific features and provides insights for designing biologically inspired architectures and brain-machine interfaces that align with neural geometry.

Jennifer Brooks How does the manifold perspective inform brain-machine interface design?

Dr. Krishna Shenoy Brain-machine interfaces decode neural activity to control external devices. Traditional approaches map individual neuron firing rates to device commands. The manifold perspective suggests instead decoding population trajectories along task-relevant manifold dimensions. This provides more robust, stable decoding because manifold structure is preserved despite neuron-level instability. We can identify manifold dimensions aligned with movement parameters and decode along these dimensions. During learning, users discover how to navigate neural manifolds to generate desired device movements, exploiting the existing manifold structure rather than creating arbitrary neural patterns. This accelerates learning and improves performance. Understanding manifold geometry also reveals which population states are accessible through voluntary control and which are constrained by circuit architecture, guiding interface design and user training strategies.

Adam Ramirez What are the main limitations of the manifold framework?

Dr. Krishna Shenoy The manifold hypothesis assumes population activity is low-dimensional, which may not hold universally—some computations might require high-dimensional representations. Estimating manifolds requires large datasets and many simultaneously recorded neurons, limiting application to small or sparse recordings. Different analysis methods can yield different manifold characterizations, making interpretation ambiguous. The relationship between manifold geometry and specific computations often remains qualitative rather than mechanistically precise—saying a manifold is curved doesn't fully specify what computation is performed. Manifolds are often characterized during restricted behavioral tasks, potentially missing structure revealed in naturalistic conditions. The framework emphasizes geometry but may underemphasize temporal dynamics or cell-type-specific contributions. Advancing the framework requires better methods for manifold estimation, clearer mappings between geometric properties and computational operations, and experimental tests manipulating manifold structure.

Jennifer Brooks How do different brain regions coordinate their manifold dynamics during behavior?

Dr. Krishna Shenoy Multi-area recordings reveal that regions have distinct manifold structures but coordinate their dynamics during behavior. During movement, premotor cortex manifolds lead motor cortex manifolds—population trajectories in premotor cortex precede and predict subsequent motor cortex trajectories. Sensory and motor manifolds show coupled dynamics during sensorimotor transformations, with sensory manifold rotations driving motor manifold state changes. Communication between regions may occur along specific manifold dimensions—subspaces aligned across areas. This suggests that inter-area coordination involves orchestrating population trajectories rather than neuron-by-neuron correspondence. Understanding multi-area computation requires characterizing how manifold structures across regions align, how information flows along cross-area dimensions, and how feedback connections modify manifold dynamics. This is analogous to understanding computation in multi-layer artificial networks through activation geometry across layers.

Adam Ramirez Can we experimentally test whether manifold geometry is computationally necessary by perturbing it?

Dr. Krishna Shenoy This is challenging but critical for establishing causality. We can use targeted optogenetic stimulation to push population states off the manifold and measure behavioral consequences and recovery dynamics. If manifolds are computationally essential, off-manifold perturbations should impair behavior more than on-manifold perturbations of similar magnitude. We can attempt to rotate or compress manifolds through sustained stimulation and test whether behavior degrades. During learning, we can constrain neural activity to remain on manifolds and test whether this accelerates or impairs skill acquisition. In brain-machine interfaces, we can design tasks requiring users to generate activity outside typical manifolds and measure feasibility. These experiments distinguish whether manifolds reflect computational necessity, optimization under constraints, or epiphenomenal consequences of circuit architecture. Interpreting results requires careful controls for stimulation artifacts and comprehensive behavioral assessment.

Jennifer Brooks How does noise interact with manifold structure? Does noise push activity off manifolds or remain constrained?

Dr. Krishna Shenoy Neural variability shows structure that respects manifolds. Noise isn't uniformly distributed across all possible neural dimensions but concentrates along certain directions. Some variability occurs tangent to the manifold—moving along the low-dimensional surface—while less occurs perpendicular—pushing off the manifold into unused dimensions. This suggests circuit dynamics actively constrain fluctuations to the manifold through recurrent connectivity and feedback. Noise along the manifold may reflect uncertainty in task-relevant variables or exploration during learning. Noise perpendicular to manifolds may be suppressed by lateral inhibition or homeostatic mechanisms. Understanding noise geometry helps distinguish signal from noise, improve decoding algorithms that leverage noise structure, and reveal how circuits balance flexibility for learning with stability for reliable computation. Noise geometry also constrains plasticity—learning rules operating on manifolds may differ from those assuming independent noise across neurons.

Adam Ramirez What role does manifold structure play in neural coding—is the geometry itself the code or a constraint on coding?

Dr. Krishna Shenoy Manifold geometry can be viewed as both code and constraint. The geometry encodes information—task variables map to manifold coordinates, relationships between variables correspond to geometric relationships. From this perspective, the brain performs computation through geometric transformations of population states. However, manifolds also constrain which codes are implementable—circuit architecture determines what manifolds are accessible, limiting possible population states. Not every mapping from variables to neural activity is realizable given architectural constraints. This dual role suggests that understanding neural coding requires characterizing both the geometric properties that encode information and the architectural constraints that generate those properties. The relationship parallels how programming languages provide both expressive power—what computations can be written—and constraints—what forms those computations must take given language syntax.

Jennifer Brooks How do manifolds change during different brain states—sleep, arousal, attention?

Dr. Krishna Shenoy Brain state profoundly affects manifold structure. During sleep, cortical manifolds show different dimensionality and geometry than waking, potentially reflecting memory consolidation operations replaying and reorganizing experience. Arousal and neuromodulatory state change manifold geometry—norepinephrine and acetylcholine can expand manifold dimensionality, increasing the space of accessible states for flexible behavior. Attention appears to rotate manifolds, aligning task-relevant dimensions with readout directions while suppressing irrelevant dimensions. Anesthesia often collapses manifold dimensionality, consistent with loss of consciousness corresponding to reduced integration. These state-dependent changes suggest that manifold geometry isn't fixed but dynamically regulated to support different cognitive modes. Understanding these changes requires characterizing manifolds across states and identifying neuromodulatory and circuit mechanisms that reshape geometry. This has implications for understanding consciousness, attention, and learning through manifold dynamics.

Adam Ramirez What are the key open questions about neural manifolds?

Dr. Krishna Shenoy What circuit mechanisms generate and maintain specific manifold geometries—can we write down dynamics equations that produce observed manifolds from known connectivity? How do manifolds adapt during learning—what plasticity rules reshape geometry to support new computations? Do different manifold geometries correspond to fundamentally different computational operations or different implementations of similar computations? How are manifolds coordinated across brain regions—do they align through shared inputs, reciprocal connections, or other mechanisms? Can we develop causal methods to test whether specific geometric properties are necessary for computation rather than correlative? How should we interpret manifolds in naturalistic behavior where task structure is less clear than laboratory settings? What is the relationship between manifold geometry and energy efficiency, robustness, or other biological constraints? Answering these requires integrating measurement, perturbation, theory, and modeling.

Jennifer Brooks The manifold framework provides geometric language for population computation but leaves open how geometry maps to mechanism.

Dr. Krishna Shenoy Precisely. We're moving from describing what populations do geometrically to understanding why circuits generate those geometries and how geometry implements specific computations.

Adam Ramirez Dr. Shenoy, thank you for clarifying how population geometry structures neural computation and what questions remain about the mechanistic basis of manifolds.

Dr. Krishna Shenoy Thank you. Understanding neural manifolds requires bridging geometry, dynamics, and circuit architecture.

Adam Ramirez That's our program for tonight. Until tomorrow, stay rigorous.

Jennifer Brooks And keep questioning. Good night.