Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representation

Abstract

It is widely believed that learning good representations is one of the main reasons for the success of deep neural networks. Although highly intuitive, there is a lack of theory and systematic approach quantitatively characterizing what representations do deep neural networks learn. In this work, we move a tiny step towards a theory and better understanding of the representations. Specifically, we study a simpler problem: How similar are the representations learned by two networks with identical architecture but trained from different initializations. We develop a rigorous theory based on the neuron activation subspace match model. The theory gives a complete characterization of the structure of neuron activation subspace matches, where the core concepts are maximum match and simple match which describe the overall and the finest similarity between sets of neurons in two networks respectively. We also propose efficient algorithms to find the maximum match and simple matches. Finally, we conduct extensive experiments using our algorithms. Experimental results suggest that, surprisingly, representations learned by the same convolutional layers of networks trained from different initializations are not as similar as prevalently expected, at least in terms of subspace match.

arXiv comment17 pages, 6 figures

References in corpus (2)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Convergent Learning: Do different neural networks learn the same representations?

Cited by in corpus (14)

• How deep is deep enough? -- Quantifying class separability in the hidden layers of deep neural networks
• Confidence Calibration for Convolutional Neural Networks Using Structured Dropout
• Emerging Cross-lingual Structure in Pretrained Language Models
• Blackbox meets blackbox: Representational Similarity and Stability Analysis of Neural Language Models and Brains
• Optimizing Mode Connectivity via Neuron Alignment
• Let's Agree to Agree: Neural Networks Share Classification Order on Real Datasets
• On Geometric Structure of Activation Spaces in Neural Networks
• CoReS: Compatible Representations via Stationarity
• Representation Transfer by Optimal Transport
• Trivial or impossible -- dichotomous data difficulty masks model differences (on ImageNet and beyond)
• ImageNet Pre-training also Transfers Non-Robustness
• Understanding the Dynamics of DNNs Using Graph Modularity
• A Methodology for Exploring Deep Convolutional Features in Relation to Hand-Crafted Features with an Application to Music Audio Modeling
• Non-uniqueness phenomenon of object representation in modelling IT cortex by deep convolutional neural network (DCNN)