Truth or Backpropaganda? An Empirical Investigation of Deep Learning Theory

Micah Goldblum, Jonas Geiping, Avi Schwarzschild, Michael Moeller, Tom Goldstein

Introduction

In contemporary artificial intelligence, a significant gap exists between theoretical guarantees and empirical performance. Practitioners frequently utilize architectures like ResNets, batch normalization, and Adam optimization because they demonstrate superior performance, despite the fact that our mathematical understanding of these components often relies on simplified "toy" models. In their ICLR 2020 paper, "Truth or Backpropaganda? An Empirical Investigation of Deep Learning Theory," Micah Goldblum and colleagues systematically test four foundational pillars of deep learning theory to determine if they hold true for modern, realistic architectures.

Methodology: A Theory-to-Experiment Pipeline

The researchers employed a rigorous "theory-to-experiment" approach. They identified specific theoretical claims, isolated the conditions under which these claims are mathematically proven, and then attempted to falsify them using realistic settings and modern architectures. Their investigation focused on four key areas: the existence of suboptimal local minima, the necessity of small-norm solutions for generalization, the stability of the Neural Tangent Kernel (NTK), and the impact of weight matrix rank on performance and robustness.

Key Findings

1. Suboptimal Local Minima are Real

A common theoretical belief is that in sufficiently large networks, all local minima are globally optimal. The authors demonstrate that this "myth" is largely a byproduct of specific initializations and optimizers. By utilizing "bad" initializations—characterized by high bias or high variance—they successfully trapped networks in suboptimal local minima with high training loss, mimicking poor linear classifiers.

2. Small Norms Aren't Always Optimal

Regularization theory often posits that keeping weight norms small is crucial for generalization. To test this, the authors introduced norm-bias, a regularizer that biases weights toward a specific, non-zero norm ($\mu$) rather than zero. Surprisingly, on CIFAR-10, a ResNet-18 using norm-bias achieved 94.86% accuracy, outperforming standard weight decay (94.06%). This suggests that the benefits of weight decay may stem from optimization dynamics rather than simply maintaining small weights.

3. The NTK Reversal in ResNets

The Neural Tangent Kernel (NTK) describes how a network's outputs evolve during training. Theory predicts that for infinite-width networks, the NTK remains constant. While this holds for simple MLPs, the authors found that modern architectures like WideResNets behave in the exact opposite manner.

Correlation coefficient between initial and final NTK across different architectures.

As illustrated in the figure above, the correlation between the initial and final NTK approaches 1.0 for MLPs as width increases. However, for WideResNets (the green line), this correlation drops significantly, falling below 0.4 as the number of parameters grows.

Relative norm of the change in NTK during training.

The relative norm of the change in the NTK further confirms this instability. In ResNets, the kernel becomes increasingly unstable as width increases, a phenomenon the authors attribute to the interaction between skip connections and batch normalization.

4. Rank is Not a Primary Driver of Generalization

While some theories suggest that low-rank weight matrices improve generalization or robustness, empirical evidence suggests otherwise. In adversarial training scenarios ($\epsilon=8/255$), high-rank networks (RankMax) achieved 35.55% robust accuracy, outperforming their low-rank counterparts (RankMin) at 33.68%.

Conclusion

The findings of Goldblum et al. highlight a critical need for a new generation of deep learning theory. Current "wide-network" theories fail to describe the models used in production today, primarily because they do not account for essential architectural features like skip connections and batch normalization. The success of ResNets, despite their highly unstable kernels, suggests that NTK stability is not a prerequisite for effective deep learning, challenging theorists to develop models that reflect the practical complexities of modern AI.

Summary updated

BibTeX
@inproceedings{goldblum2020truth,
  author = {Micah Goldblum and Jonas Geiping and Avi Schwarzschild and Michael Moeller and Tom Goldstein},
  title = {{Truth or Backpropaganda? An Empirical Investigation of Deep Learning Theory}},
  booktitle = {International Conference on Learning Representations},
  year = {2020},
  url = {https://openreview.net/forum?id=LidbmSbyrxs},
}
Figure