[Paper Review] On the emergence of tetrahedral symmetry in the final and penultimate layers of neural network classifiers.

E Weinan, Stephan Wojtowytsch|arXiv (Cornell University)|Dec 10, 2020

Neural Networks and Applications11 references9 citations

TL;DR

This paper analytically explains the observed tetrahedral symmetry in the penultimate layer of deep neural network classifiers, showing that class-specific data points converge to vertices of a regular simplex in high-dimensional space due to the network's expressive capacity. It further demonstrates that in shallow networks, such symmetry does not emerge in the final layer, highlighting the critical role of deeper architecture in shaping geometric data configurations.

ABSTRACT

A recent numerical study observed that neural network classifiers enjoy a large degree of symmetry in the penultimate layer. Namely, if $h(x) = Af(x) +b$ where $A$ is a linear map and $f$ is the output of the penultimate layer of the network (after activation), then all data points $x_{i, 1}, \dots, x_{i, N_i}$ in a class $C_i$ are mapped to a single point $y_i$ by $f$ and the points $y_i$ are located at the vertices of a regular $k-1$-dimensional standard simplex in a high-dimensional Euclidean space. We explain this observation analytically in toy models for highly expressive deep neural networks. In complementary examples, we demonstrate rigorously that even the final output of the classifier $h$ is not uniform over data samples from a class $C_i$ if $h$ is a shallow network (or if the deeper layers do not bring the data samples into a convenient geometric configuration).

Motivation & Objective

• To explain the empirical observation of symmetric data clustering in the penultimate layer of deep neural network classifiers.
• To investigate why such symmetry emerges in deep networks but not in shallow ones.
• To analyze the geometric configuration of class representations in both penultimate and final layers of neural networks.
• To establish conditions under which data points from the same class are mapped to symmetric positions in high-dimensional space.

Proposed method

• Analytical modeling of highly expressive deep neural networks using linear transformations and nonlinear activations.
• Defining the penultimate layer output as $ h(x) = Af(x) + b $, where $ f $ is the activation output before the final classifier layer.
• Using geometric arguments to show that data points from the same class $ C_i $ are mapped to a single point $ y_i $, forming a regular $ k-1 $-dimensional simplex.
• Constructing complementary examples with shallow networks to demonstrate lack of symmetry in the final output layer.
• Applying symmetry and invariance principles to characterize the emergence of simplex configurations in high-dimensional space.

Experimental results

Research questions

• RQ1 Why do data points from the same class cluster at the vertices of a regular simplex in the penultimate layer of deep neural networks?
• RQ2 Under what architectural conditions does tetrahedral symmetry emerge in the representation space of neural network classifiers?
• RQ3 Why does such symmetry fail to appear in the final output layer when the network is shallow?
• RQ4 How do deeper layers contribute to the geometric organization of class representations in the network?

Key findings

• In highly expressive deep networks, data points from the same class are mapped to a single point in the penultimate layer, forming a regular simplex configuration in high-dimensional space.
• The symmetry arises due to the interplay between the network's depth, expressivity, and the linear transformation applied before the final classification layer.
• In shallow networks, the final classifier output $ h $ does not exhibit uniformity over samples from the same class, indicating absence of symmetric clustering.
• The geometric configuration of class representations is contingent on the presence of deeper layers that reorganize data into favorable manifolds.