Do Wide and Deep Networks Learn the Same Things? Uncovering How Neural Network Representations Vary with Width and Depth

Abstract

A key factor in the success of deep neural networks is the ability to scale models to improve performance by varying the architecture depth and width. This simple property of neural network design has resulted in highly effective architectures for a variety of tasks. Nevertheless, there is limited understanding of effects of depth and width on the learned representations. In this paper, we study this fundamental question. We begin by investigating how varying depth and width affects model hidden representations, finding a characteristic block structure in the hidden representations of larger capacity (wider or deeper) models. We demonstrate that this block structure arises when model capacity is large relative to the size of the training set, and is indicative of the underlying layers preserving and propagating the dominant principal component of their representations. This discovery has important ramifications for features learned by different models, namely, representations outside the block structure are often similar across architectures with varying widths and depths, but the block structure is unique to each model. We analyze the output predictions of different model architectures, finding that even when the overall accuracy is similar, wide and deep models exhibit distinctive error patterns and variations across classes.

arXiv commentICLR 2021

References in corpus (19)

• Neural Tangent Kernel: Convergence and Generalization in Neural Networks
• Deep Neural Networks as Gaussian Processes
• Similarity of Neural Network Representations Revisited
• Algorithms for Learning Kernels Based on Centered Alignment
• On Lazy Training in Differentiable Programming
• SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability
• Bayesian Deep Convolutional Networks with Many Channels are Gaussian Processes
• ResNet with one-neuron hidden layers is a Universal Approximator
• Approximating Continuous Functions by ReLU Nets of Minimal Width
• Representation Benefits of Deep Feedforward Networks
• A Closer Look at Deep Learning Heuristics: Learning rate restarts, Warmup and Distillation
• The large learning rate phase of deep learning: the catapult mechanism
• What shapes feature representations? Exploring datasets, architectures, and training
• The effect of task and training on intermediate representations in convolutional neural networks revealed with modified RV similarity analysis
• Learning From Brains How to Regularize Machines
• Deep neuroethology of a virtual rodent
• Probing the State of the Art: A Critical Look at Visual Representation Evaluation
• Investigating Multilingual NMT Representations at Scale
• Post Selection Inference with Kernels

Cited by in corpus (12)

• Do Vision Transformers See Like Convolutional Neural Networks?
• On Representation Knowledge Distillation for Graph Neural Networks
• Gaussian RBF Centered Kernel Alignment (CKA) in the Large Bandwidth Limit
• Zen-NAS: A Zero-Shot NAS for High-Performance Deep Image Recognition
• Why Adversarial Reprogramming Works, When It Fails, and How to Tell the Difference
• GraphITE: Estimating Individual Effects of Graph-structured Treatments
• Trivial or impossible -- dichotomous data difficulty masks model differences (on ImageNet and beyond)
• Properties of the After Kernel
• Probing neural networks with t-SNE, class-specific projections and a guided tour
• Why Do Better Loss Functions Lead to Less Transferable Features?
• Understanding the Dynamics of DNNs Using Graph Modularity
• Analyze and Design Network Architectures by Recursion Formulas