Abstract
Brains and artificial neural networks compute with continuous variables such as object position or stimulus orientation. However, the complex variability in neural responses makes it difficult to link internal representational structure to task performance. We develop a statistical-mechanical theory of regression capacity that relates linear decoding efficiency of continuous variables to geometric properties of neural manifolds. Our theory handles complex neural variability and applies to real data, revealing increasing capacity for decoding object position and size along the monkey visual stream.
Introduction
In both biological and artificial intelligence, neural population activity is often conceptualized as existing on "manifolds"—structured surfaces within a high-dimensional state space. While robust theories have been established to explain how the geometry of these manifolds influences a system's ability to categorize objects (e.g., distinguishing a cat from a dog), a general theory for continuous variables (e.g., determining the precise orientation or position of that cat) has remained elusive.
Existing regression models frequently rely on rigid parametric assumptions that fail to account for the complex "nuisance" variability inherent in real-world neural data. This paper addresses this gap by introducing the concept of regression capacity: a quantitative measure of the maximum information a neural population can store such that a linear decoder can still retrieve a continuous variable within a specified error tolerance.
The Geometry of Regression
A central innovation of this work is the definition of $\epsilon$-regressibility. A neural manifold is considered regressible if all its constituent activity patterns lie within a narrow "slab" of width $2\epsilon$ centered around a linear decoding hyperplane. This geometric interpretation allows the researchers to link decoding performance directly to the physical shape of the neural representation.
The theory demonstrates that regression capacity is fundamentally constrained by the "size" of neural noise. Specifically, capacity increases as the intrinsic dimensionality ($D$) and radius ($R$) of the neural manifolds decrease. In essence, smaller, flatter manifolds are significantly easier for a downstream neuron to "read" accurately.
Theoretical Frameworks
The authors employ two complementary theoretical approaches to quantify these relationships:
- Mean-Field Theory: Utilizing tools from statistical mechanics (specifically the Gardner volume approach), they derive analytical formulas for the capacity of synthetic manifold models. This reveals how geometric parameters dictate decodability in the "thermodynamic limit," where the number of neurons and stimuli are very large.
- Instance-Based Theory: For empirical datasets where the underlying generative model is unknown, they developed a non-parametric estimator. This method identifies the "critical dimension"—the minimum number of neurons required to achieve a target decoding accuracy—using random projections and quadratic programming.
Results and Biological Implications
The study validated these theories using both synthetic manifolds and electrophysiological recordings from the macaque ventral stream (areas V4 and IT). The task involved decoding object pose parameters—size and position—against complex, variable backgrounds.
Universal Rescaling
A key finding is the discovery of "universal rescaling." The authors found that correlations between different stimuli or labels do not fundamentally alter the nature of the regression problem; instead, they effectively rescale the "equivalent tolerance." This suggests a mathematically consistent strategy within the brain for handling correlated noise across diverse types of variability.
Hierarchical Refinement
The theory provides a formal geometric explanation for the "untangling" of variables in the visual stream. As information progresses from raw pixels to area V4 and finally to IT, the manifolds are reshaped to become more linearly regressible. Area IT exhibited significantly higher regression capacity than V4, even when nuisance background variability remained high. This progression matches traditional cross-validated decoding results but offers a more direct, geometric interpretation: the capacity measure represents the actual number of neurons a downstream "reader" requires to achieve a target accuracy $\epsilon$.
Discussion
While this framework provides a powerful tool for understanding linear readouts—a standard benchmark for what a single downstream neuron can perceive—it is currently limited to linear decoders. Future research may extend these principles to the non-linear decoders common in artificial systems. Additionally, applying this geometric lens to motor control and navigation tasks could reveal how the brain manages continuous variables like joint angles or spatial coordinates during active behavior.
Summary updated