How Do Training Methods Influence the Utilization of Vision Models?

TL;DR

The findings reveal that the training method strongly influences which layers become critical to the decision function for a given task, and improved training regimes and self-supervised training increase the importance of early layers while significantly under-utilizing deeper layers.

Abstract

Not all learnable parameters (e.g., weights) contribute equally to a neural network's decision function. In fact, entire layers' parameters can sometimes be reset to random values with little to no impact on the model's decisions. We revisit earlier studies that examined how architecture and task complexity influence this phenomenon and ask: is this phenomenon also affected by how we train the model? We conducted experimental evaluations on a diverse set of ImageNet-1k classification models to explore this, keeping the architecture and training data constant but varying the training pipeline. Our findings reveal that the training method strongly influences which layers become critical to the decision function for a given task. For example, improved training regimes and self-supervised training increase the importance of early layers while significantly under-utilizing deeper layers. In contrast, methods such as adversarial training display an opposite trend. Our preliminary results extend previous findings, offering a more nuanced understanding of the inner mechanics of neural networks. Code: https://github.com/paulgavrikov/layer_criticality

Figures (5)

• Figure 1: Training methods determine what layers become critical. We measure the criticality of fifty different ResNet-50-based models that all utilize the same exact network architecture and training data (ImageNet-1k) but differ in their training methods. Darker spots denote layers that are critical, i.e., in significantly different predictions and decreased performance after reset. Brighter spots are auxiliary, i.e., resetting these layers does not significantly affect the model. We denote the average (mean$\pm$std) layer criticality for both, a model across layers on the right, for a layer across model on the bottom.
• Figure 2: Adversarial training increases the average criticality proportional to the training attack budget $\epsilon$. We ablate $\ell_\infty$ from $\ell_2$-norm training but do not observe any significant differences in their trends. The marker size in the plot indicates the validation accuracy on ImageNet-1k (larger is better).
• Figure 3: Correlation between average network criticality and performance on ImageNet-1k.
• Figure 4: Criticality difference to the baseline. In addition to the plot in \ref{['fig:weight_reset_r50_training']}, we here show the difference to the baseline model resnet. Positive numbers indicate increases in criticality and negative numbers decrease.
• Figure 5: Criticality standard error. In addition to the plot in \ref{['fig:weight_reset_r50_training']}, we here show the standard error in criticality measurements over 3 runs with different seeds for the randomization.