Abstract
Recently, growth in our understanding of the computations performed in both biological and artificial neural networks has largely been driven by either low-level mechanistic studies or global normative approaches. However, concrete methodologies for bridging the gap between these levels of abstraction remain elusive. In this work, we investigate the internal mechanisms of neural networks through the lens of neural population geometry, aiming to provide understanding at an intermediate level of abstraction, as a way to bridge that gap. Utilizing manifold capacity theory (MCT) from statistical physics and manifold alignment analysis (MAA) from high-dimensional statistics, we probe the underlying organization of task-dependent manifolds in deep neural networks and macaque neural recordings. Specifically, we quantitatively characterize how different learning objectives lead to differences in the organizational strategies of these models and demonstrate how these geometric analyses are connected to the decodability of task-relevant information. These analyses present a strong direction for bridging mechanistic and normative theories in neural networks through neural population geometry, potentially opening up many future research avenues in both machine learning and neuroscience.
Introduction
Deep neural networks (DNNs) have achieved remarkable success, yet their internal representational strategies often remain a "black box." While simple accuracy metrics can compare models, they fail to explain the structural differences in how information is organized. This research aims to bridge the gap between low-level mechanistic views (individual neurons) and high-level normative theories (task performance) by analyzing neural population geometry. By treating neural activity as high-dimensional shapes, or "manifolds," researchers can distinguish between models that merely "cheat" by over-specializing and those that capture the true underlying structure of the world, much like the biological brain.
Geometric Framework and Methodology
The study introduces a dual-lens approach to characterize neural manifolds:
- Manifold Capacity Theory (MCT): Grounded in statistical physics, MCT quantifies the "separability" of object categories. It measures the Radius (size), Dimension (complexity), and Capacity (linear classifiability) of neural manifolds.
- Manifold Alignment Analysis (MAA): This novel contribution examines the relative orientation of manifolds. Using metrics like Bures covariance distance and Signal Mismatch Distance, MAA assesses how well a network "lines up" its representations across different objects to efficiently use its representational space.
The researchers compared ResNet-50 models trained with various objectives—supervised learning versus unsupervised methods like SimCLR and Barlow Twins—against neural recordings from the macaque visual cortex (V4 and IT).
Datasets
The analysis utilized three key data sources:
- Synthetic 3D Objects: A controlled dataset of 64 objects across 8 superclasses, allowing for precise manipulation of six latent parameters: size, position (x, y), and rotation (x, y, z).
- ImageNet: Used for pre-training the DNN models.
- Macaque Neural Recordings: Multi-electrode array data from the V4 and IT cortex of monkeys viewing the synthetic 3D stimuli.
Key Insights and Results
The geometric analysis revealed fundamental differences in how supervised and unsupervised models organize information:
- Specialization vs. Generalization: Supervised models are "overspecialized." They achieve high accuracy by aggressively shrinking object manifolds (reducing Radius and Dimension). In contrast, unsupervised models and the macaque brain maintain larger, more complex manifolds, preserving general information about the stimuli.
- The Alignment Strategy: Unsupervised models compensate for larger manifolds through alignment. By aligning the axes of variation across different objects, these networks use their representational space more efficiently, which is quantitatively linked to better performance in decoding latent features (regression).
- Biological Plausibility: Unsupervised models showed a significantly higher correlation with macaque IT cortex data ($r \approx 0.8$) compared to supervised models. The macaque brain exhibits "lazy" compression, aligning more closely with unsupervised organizational strategies.
- Beyond RSA: Traditional Representational Similarity Analysis (RSA) often fails to detect these structural nuances. Neural population geometry provides a higher-resolution view of the internal strategies employed by different learning objectives.
Conclusion and Future Directions
This work demonstrates that neural population geometry is a powerful tool for bridging mechanistic and normative theories. While the link between geometry and classification (MCT) is well-established, this study takes a significant step toward a formal theory for regression manifolds via MAA.
Future research could extend these geometric signatures to newer architectures like Vision Transformers (ViTs) or other sensory modalities. Ultimately, these geometric insights could serve as "regularizers" to design AI systems that are inherently more robust, efficient, and brain-like.
Summary updated