The learning phases in NN: From Fitting the Majority to Fitting a Few

Johannes SchneiderarXiv 2022 · arXiv

Introduction

How do deep neural networks (DNNs) learn? For years, the Information Bottleneck (IB) theory has been the dominant explanation, suggesting that networks first "fit" the data and then "compress" it by discarding irrelevant information. However, the IB theory is mathematically elusive and its empirical validity is often debated.

This paper proposes an alternative, more intuitive framework: a two-phase learning process consisting of Prototyping followed by Fitting the Few. By monitoring how a network's internal representations evolve, the authors demonstrate that networks don't just compress; they actively distort general class prototypes to accommodate difficult outliers.

Methodology: The Dual-Lens Approach

To track learning dynamics without relying on abstract mutual information metrics, the study employs two tangible signals:

  1. Linear Separability ($Acc^{(t)}$): By training a simple linear classifier on frozen hidden layer activations, the authors measure how well the network has organized the data for classification.
  2. Reconstruction Error ($Rec^{(t)}$): By training a separate decoder to recreate the original input from hidden activations, they measure how much information about the input is preserved.

This "dual-lens" approach was tested across standard benchmarks like MNIST, Fashion-MNIST, and CIFAR-10/100 using architectures such as VGG-11 and ResNet-10.

Phase 1: Prototyping (Fitting the Majority)

Early in training, the network enters the "Prototyping" phase. Here, the weights begin to resemble the average of the classes they are learning to identify. In Fashion-MNIST, for instance, a weight matrix might visually manifest as a ghostly T-shirt or trouser. During this stage, the network captures the most prominent features shared by the majority of samples in a class. This leads to a rapid increase in classification accuracy and a corresponding decrease in reconstruction error, as the "prototype" serves as an efficient, low-noise representation of the data.

Visual evolution of class representations across training iterations, showing the emergence and eventual distortion of prototypes.

Phase 2: Fitting the Few (The U-Shaped Curve)

The most striking finding of this research is that reconstruction error eventually stops decreasing and begins to rise, even as classification accuracy continues to improve. This creates a distinct "U-shaped" curve for reconstruction loss.

The authors attribute this to the "Fitting the Few" phase. Once the network has mastered the "easy" majority samples, gradient descent forces it to minimize the loss on "hard" samples—the outliers. To classify these exceptions, the network must shift its weights away from the clean, general prototype. This process introduces "noise" or high-frequency distortions into the hidden representations. While this helps the network classify the outliers, it makes the original input much harder to reconstruct from those same activations.

The U-shaped reconstruction loss curve alongside rising classification accuracy, demonstrating the transition between learning phases.

Key Insights and Theoretical Support

  • Reinterpreting Compression: What was previously interpreted as "compression" (discarding information) may actually be a byproduct of "distortion." The network isn't necessarily trying to be efficient; it is sacrificing a general representation to satisfy the loss requirements of difficult edge cases.
  • Mathematical Foundation: The authors provide a theoretical model for single-layer networks, proving that reconstruction error is guaranteed to follow this U-shape when the network transitions from learning class averages to correcting errors on individual noisy samples.

Conclusion

The "Fitting the Majority to Fitting a Few" model provides a clear, visual, and mathematically grounded alternative to the Information Bottleneck theory. It suggests that the "noise" seen in late-stage neural network weights is not random, but a functional necessity for handling the complexity of real-world data distributions. Future work may explore whether this transition is a universal property of all gradient-based learning systems, potentially offering new ways to regularize models or detect overfitting.

Summary updated

Abstract

The learning dynamics of deep neural networks are subject to controversy. Using the information bottleneck (IB) theory separate fitting and compression phases have been put forward but have since been heavily debated. We approach learning dynamics by analyzing a layer's reconstruction ability of the input and prediction performance based on the evolution of parameters during training. We show that a prototyping phase decreasing reconstruction loss initially, followed by reducing classification loss of a few samples, which increases reconstruction loss, exists under mild assumptions on the data. Aside from providing a mathematical analysis of single layer classification networks, we also assess the behavior using common datasets and architectures from computer vision such as ResNet and VGG.
BibTeX
@misc{schneider2022learning,
  author = {Johannes Schneider},
  title = {{The learning phases in NN: From Fitting the Majority to Fitting a Few}},
  year = {2022},
  eprint = {2202.08299},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2202.08299},
}
Figure