Separability and Geometry of Object Manifolds in Deep Neural Networks

Uri Cohen, SueYeon Chung, Daniel D. Lee, Haim Sompolinsky

Abstract: Stimuli are represented in the brain by the collective population responses of sensory neurons, and an object presented under varying conditions gives rise to a collection of neural population responses called an object manifold. Changes in the object representation along a hierarchical sensory system are associated with changes in the geometry of those manifolds, and recent theoretical progress connects this geometry with classification capacity, a quantitative measure of the ability to support object classification. Deep neural networks trained on object classification tasks are a natural testbed for the applicability of this relation. We show how classification capacity improves along the hierarchies of deep neural networks with different architectures. We demonstrate that changes in the geometry of the associated object manifolds underlie this improved capacity, and shed light on the functional roles different levels in the hierarchy play to achieve it, through orchestrated reduction of manifolds’ radius, dimensionality and inter-manifold correlations.

Introduction: Beyond the "Untangling" Metaphor

The "untangling" hypothesis has long been a popular intuition in both neuroscience and machine learning. It suggests that sensory systems transform messy, overlapping inputs—where different objects might share similar pixel-level features—into clean, linearly separable representations. However, "untangling" has historically remained a vague metaphor.

Conceptual diagram of object manifolds being 'untangled' through successive layers of a neural hierarchy.

The challenge is that the separability of objects depends on a complex interplay of factors: the number of neurons, the number of objects, their shapes, their sizes, and their mutual correlations. This paper seeks to replace the metaphor with rigorous mathematics, addressing a fundamental question: Which specific geometric properties of internal representations actually change during learning to make classification possible?

Theoretical Framework and Methods

The authors utilize a theoretical framework called the Mean-Field Theory of Manifold Classification Capacity. This framework treats an "object" not as a single point in neural space, but as a manifold—a collection of points representing the same object under varying conditions, such as different viewing angles or lighting.

To quantify the transformation of these manifolds, the authors define several key metrics:

  • Classification Capacity ($\alpha_c$): The core metric, measuring the maximum number of object manifolds that can be linearly separated per neuron.
  • Manifold Radius ($R_M$): The effective "size" or spatial extent of the manifold.
  • Manifold Dimension ($D_M$): The effective number of axes along which the manifold spreads.
  • Anchor Points: A generalization of "support vectors" for manifolds; these are the points on the manifold's edge that determine the position of the separating hyperplane.

The researchers analyzed both point-cloud manifolds from ImageNet and synthetic smooth manifolds created via affine transformations. To isolate the effects of learning, they compared fully trained Deep Convolutional Neural Networks (DCNNs) like AlexNet and VGG-16 against untrained baselines with random weights.

Key Insights: Orchestrated Geometric Reduction

The study reveals that the improvement in classification capacity is not a byproduct of architecture alone, but an emergent property of the learning process. This improvement is driven by an "orchestrated reduction" in three areas:

  1. Size: Manifolds physically shrink (radius decreases).
  2. Flatness: Manifolds become more compressed (dimension decreases).
  3. Separation: The centers of different manifolds move away from each other (correlations decrease).

Interestingly, the authors found a trade-off within the network's building blocks. For instance, ReLU nonlinearities actually increase manifold dimension—making them more complex—but compensate by sparsifying the representation and reducing correlations. Conversely, Pooling layers reduce radius and dimension but increase correlations. It is only through the sequential combination of these operations that the network achieves a net gain in capacity.

Key Results

The empirical results across various architectures (AlexNet, VGG, ResNet) show consistent trends that support the theoretical model. In AlexNet, the relative classification capacity increases by nearly an order of magnitude from the input layer to the final feature layer.

Quantitative evidence showing the increase in classification capacity (alpha_c) and the decrease in manifold dimension (D_M) across network layers.

As shown in the data, the effective manifold dimension ($D_M$) drops dramatically deeper in the network. For point-cloud manifolds, $D_M$ falls from over 80 in early layers to roughly 20 in the final feature layer. Simultaneously, center correlations between different objects decay, meaning the network learns to push different categories into distinct, non-overlapping regions of the high-dimensional feature space.

Limitations and Future Directions

While this work provides a powerful geometric lens for understanding DNNs, several questions remain. The current theory assumes deterministic neural responses, whereas biological brains are inherently noisy. Future research will need to incorporate trial-to-trial variability to better model biological systems. Additionally, while the authors suggest these geometric measures relate to generalization, a formal mathematical link between manifold geometry and test-set error remains to be explicitly proven. Finally, it is yet to be determined if this "untangling" process occurs similarly in networks trained for non-classification tasks, such as generative modeling or self-supervised learning.

Summary updated

Abstract

Abstract Stimuli are represented in the brain by the collective population responses of sensory neurons, and an object presented under varying conditions gives rise to a collection of neural population responses called an object manifold . Changes in the object representation along a hierarchical sensory system are associated with changes in the geometry of those manifolds, and recent theoretical progress connects this geometry with classification capacity , a quantitative measure of the ability to support object classification. Deep neural networks trained on object classification tasks are a natural testbed for the applicability of this relation. We show how classification capacity improves along the hierarchies of deep neural networks with different architectures. We demonstrate that changes in the geometry of the associated object manifolds underlie this improved capacity, and shed light on the functional roles different levels in the hierarchy play to achieve it, through orchestrated reduction of manifolds’ radius, dimensionality and inter-manifold correlations.
BibTeX
@inproceedings{cohen2020separability,
  author = {Uri Cohen and SueYeon Chung and Daniel D. Lee and Haim Sompolinsky},
  title = {{Separability and Geometry of Object Manifolds in Deep Neural Networks}},
  booktitle = {Nature Communications},
  year = {2020},
  doi = {10.1038/s41467-020-14578-5},
}
Figure