[Paper Review] Spectral Analysis of Latent Representations

Justin Shenk, Mats L. Richter+2 more|arXiv (Cornell University)|Jul 19, 2019

Advanced Neural Network Applications7 references4 citations

TL;DR

The paper introduces Layer Saturation, a fast, on-the-fly metric based on spectral analysis of latent representations to assess neural network architecture quality during training. By computing the proportion of eigenvalues explaining 99% of variance, it reveals over- or underparameterization and correlates strongly with generalization performance.

ABSTRACT

We propose a metric, Layer Saturation, defined as the proportion of the number of eigenvalues needed to explain 99% of the variance of the latent representations, for analyzing the learned representations of neural network layers. Saturation is based on spectral analysis and can be computed efficiently, making live analysis of the representations practical during training. We provide an outlook for future applications of this metric by outlining the behaviour of layer saturation in different neural architectures and problems. We further show that saturation is related to the generalization and predictive performance of neural networks.

Motivation & Objective

• To develop a fast, interpretable metric for monitoring latent representation quality during neural network training.
• To enable real-time detection of overparameterization and underparameterization without relying on validation sets.
• To provide insight into information flow and intrinsic dimensionality across layers in deep networks.
• To guide model architecture selection by linking saturation patterns to predictive performance and problem complexity.
• To offer a lightweight alternative to computationally heavy analysis tools like SVCCA, usable during training.

Proposed method

• Compute the autocovariance matrix of pre-activation latent representations using an online, incremental update rule to avoid storing full activation histories.
• Apply singular value decomposition (SVD) to the covariance matrix to extract eigenvalues and eigenvectors.
• Determine the intrinsic dimensionality as the number of top eigenvalues required to explain 99% of total variance.
• Define Layer Saturation as the ratio of intrinsic dimensionality to the total number of neurons/filters in a layer, yielding a value between 0 and 1.
• Compute model-wide average saturation by averaging layer-wise saturation across all non-output convolutional and fully connected layers.
• Use the metric in real time during training to track changes in representation complexity and detect problematic architectures early.

Experimental results

Research questions

• RQ1 How does layer-wise saturation evolve during training, and what patterns indicate over- or underparameterization?
• RQ2 Can saturation serve as a reliable proxy for generalization performance without requiring validation data?
• RQ3 How does saturation vary with architectural depth and filter count in convolutional networks?
• RQ4 Does saturation correlate with the complexity of the learning task, such as between CIFAR10 and Cat vs. Dog?
• RQ5 Can saturation be used to detect architectural flaws such as excessive depth or insufficient capacity in real time?

Key findings

• Layer saturation correlates strongly with model generalization performance, with lower saturation generally indicating better performance.
• Overparameterized networks exhibit a long-tail distribution of high saturation across deeper layers, especially in deep architectures like VGG19.
• Underparameterized networks show high saturation from the start, with a rapid drop-off in early layers, enabling early detection of poor capacity.
• The number of filters per layer has a more severe impact on saturation and performance than depth alone, particularly on complex datasets like CIFAR10.
• Average saturation across the network serves as a reliable indicator of whether the architecture is sufficiently complex for the given task, with performance degrading sharply above ~45% saturation on complex problems.
• Saturation patterns are sensitive to dataset complexity, with the same architecture showing lower saturation and longer tails on simpler tasks like Cat vs. Dog compared to CIFAR10.