On the Relationship Between Representation Geometry and Generalization in Deep Neural Networks

A study analyzing 52 pretrained ImageNet models found that effective dimension, a geometric measurement of neural network representations, predicts model accuracy with a correlation of r=0.75. This unsupervised metric outperforms traditional measures like model capacity and generalizes across computer vision and NLP tasks, establishing bidirectional causality between representation geometry and generalization performance.

On the Relationship Between Representation Geometry and Generalization in Deep Neural Networks

Unsupervised Geometric Metric Emerges as Powerful Predictor of AI Model Performance

A groundbreaking study analyzing dozens of pretrained models reveals that a simple, unsupervised geometric measurement of a neural network's internal representations is a stronger predictor of its final performance than traditional metrics like model size or capacity. The research, published in a paper on arXiv, establishes a robust, causal link between a model's representation geometry—specifically its effective dimension—and its accuracy across both computer vision and natural language processing tasks.

Analyzing Geometry Across 52 Pretrained ImageNet Models

The researchers conducted a systematic analysis of 52 pretrained models spanning 13 distinct architecture families on the ImageNet benchmark. They measured the effective dimension of the models' internal representations, a metric that quantifies how data is spread across the dimensions of a network's activation space. The results were striking: the output effective dimension alone achieved a partial correlation of r=0.75 (p < 10⁻¹⁰) with model accuracy after controlling for model capacity. Conversely, a related metric called total compression showed a strong negative correlation of r=-0.72, indicating that models which compress information less tend to perform better.

Findings Generalize to NLP and Establish Causality

The predictive power of effective dimension proved to be domain-agnostic. The findings successfully replicated on the CIFAR-10 dataset and generalized to natural language processing. For 8 encoder models on SST-2 and MNLI tasks, and 15 decoder-only large language models (LLMs) on AG News, effective dimension predicted performance with a correlation of r=0.69 (p=0.004). Critically, model size showed no significant predictive power in this context (r=0.07).

Most importantly, the study moved beyond correlation to establish bidirectional causality. Artificially degrading a model's representation geometry by injecting noise (Gaussian, Uniform, Dropout, or Salt-and-pepper) caused a predictable drop in accuracy (r=-0.94, p < 10⁻⁹). Conversely, improving the geometry by applying Principal Component Analysis (PCA) to preserve 95% of variance allowed models to maintain accuracy with a negligible average drop of just -0.03 percentage points.

Why This Discovery Matters for AI Development

This research provides a fundamental new lens for understanding, evaluating, and potentially designing neural networks.

  • Unsupervised Evaluation: Effective dimension can be computed without any task labels, offering a powerful tool for pre-deployment model assessment and selection.
  • Beyond Scale: The results challenge the predominant focus on model size as the primary lever for performance, highlighting intrinsic geometric quality as a critical factor.
  • Architecture-Agnostic Insight: The relationship holds across diverse model families, suggesting a universal principle of learning dynamics related to representation structure.
  • Causal Understanding: Establishing that geometry causally influences performance opens new avenues for model improvement through targeted geometric regularization or optimization.

By demonstrating that effective dimension provides domain-agnostic predictive and causal information, this work shifts focus toward the intrinsic geometric properties of learned representations as a key to unlocking and explaining neural network performance.

常见问题