Are All Layers Created Equal?

By Chiyuan Zhang, Samy Bengio, Yoram Singer. 2019

166 citations recorded by the source corpus.

Understanding deep neural networks has been a major research objective in recent years with notable theoretical progress. A focal point of those studies stems from the success of excessively large networks which defy the classical wisdom of uniform convergence and learnability. We study empirically the layer-wise functional structure of overparameterized deep models. We provide evidence for the heterogeneous characteristic of layers. To do so, we introduce the notion of robustness to post-training re-initialization and re-randomization. We show that the layers can be categorized as either ``ambient'' or ``critical''. Resetting the ambient layers to their initial values has no negative consequence, and in many cases they barely change throughout training. On the contrary, resetting the critical layers completely destroys the predictor and the performance drops to chanceh. Our study provides further evidence that mere parameter counting or norm accounting is too coarse in studying generalization of deep models, and flatness or robustness analysis of the models needs to respect the network architectures.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

What is being transferred in transfer learning? (2020, methodology): “In Figure 5 we analyze criticality of different modules in the same way as [45].”
Anatomy of Catastrophic Forgetting: Hidden Representations and Task Semantics (2020, methodology): “Our contributions are as follows: • We perform a thorough empirical investigation of how catastrophic forgetting affects the hidden representations of neural networks, by using representational similarity measures [21, 22], layer freezing and layer reset experiments [23].”
When BERT Plays the Lottery, All Tickets Are Winning (2020, background): “We now know that most Trans-former heads and even layers can be pruned (Voita et al., 2019; Michel et al., 2019; Kovaleva et al., 2019), but it is not clear whether that is due to redundant weights, or to some parts of the model simply being “inactive” (Zhang et al., 2019).”
How fine can fine-tuning be? Learning efficient language models (2020, background): “There is no straightforward control of the amount of weights pruned in previous reports of supermask training (Zhang et al., 2019; Mallya et al., 2018).”
Masking as an Efficient Alternative to Finetuning for Pretrained Language Models (2020, background): “Neural network layers present heterogeneous characteristics (Zhang et al., 2019) when being applied to tasks.”
Surprises in High-Dimensional Ridgeless Least Squares Interpolation (2019, unknown): “[71] Chiyuan Zhang, Samy Bengio, and Yoram Singer.”
The intriguing role of module criticality in the generalization of deep networks (2019, methodology): “For example, Figures 7a, 7b in the appendix illustrate the pattern when we changed the initialization to Fixup (Zhang et al., 2019b), and when we replace SGD with Adam (Kingma & Ba, 2014) respectively.”
Training BatchNorm and Only BatchNorm: On the Expressive Power of Random Features in CNNs (2020, methodology): “Zhang et al. (2019a) show that many individual layers in trained CNNs can be reset to their random i.i.d. initializations with little impact on accuracy.”
Backward Feature Correction: How Deep Learning Performs Deep Learning (2020, background): “The higher-level layers have diminishing contributions to the signal (see Figure 3 for an illustration and we also refer to [60] for concrete measures).”
On the geometry of generalization and memorization in deep neural networks (2021, methodology): “Whereas Zhang et al. (2019) only examine E = 0, we rewind to various E throughout training to probe the dynamics.”
A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning (2022, background): “A possible reason is that shallow layers tend to provide generalpurpose representations, whereas later layers specialize [35, 3, 4, 56, 59].”
Picking Winning Tickets Before Training by Preserving Gradient Flow (2020, background): “Moreover, it has been known that those convolutional layers at the top usually learn highly sparse features and thus more weights can be pruned (Zhang et al., 2019a).”
A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations (2023, unknown): “Zhang et al. [201] empirically divide the layers into either “ambient” or “critical”.”