Similarity of Neural Network Representations Revisited

Recent work has sought to understand the behavior of neural networks by comparing representations between layers and between different trained models. We examine methods for comparing neural network representations based on canonical correlation analysis (CCA). We show that CCA belongs to a family of statistics for measuring multivariate similarity, but that neither CCA nor any other statistic that is invariant to invertible linear transformation can measure meaningful similarities between representations of higher dimension than the number of data points. We introduce a similarity index that measures the relationship between representational similarity matrices and does not suffer from this limitation. This similarity index is equivalent to centered kernel alignment (CKA) and is also closely connected to CCA. Unlike CCA, CKA can reliably identify correspondences between representations in networks trained from different initializations.

[1]  L. Tucker.  A METHOD FOR SYNTHESIS OF FACTOR ANALYSIS STUDIES , 1951 .

[2]  H. Vinod.  Canonical ridge and econometrics of joint production , 1976 .

[3]  P. Robert,et al.  A Unifying Tool for Linear Multivariate Statistical Methods: The RV‐Coefficient , 1976 .

[4]  J. Ramsay,et al.  Matrix correlation , 1984 .

[5]  Yann LeCun,et al.  Second Order Properties of Error Surfaces: Learning Time and Generalization , 1990, NIPS 1990.

[6]  G. Golub,et al.  The canonical correlations of matrix pairs and their numerical computation , 1992 .

[7]  Robert Hecht-Nielsen,et al.  On the Geometry of Feedforward Neural Network Error Surfaces , 1993, Neural Computation.

[8]  S Edelman,et al.  Representation is representation of similarities , 1996, Behavioral and Brain Sciences.

[9]  Garrison W. Cottrell,et al.  Content and cluster analysis: Assessing representational similarity in neural systems , 2000 .

[10]  N. Cristianini,et al.  On Kernel-Target Alignment , 2001, NIPS.

[11]  A. Ishai,et al.  Distributed and Overlapping Representations of Faces and Objects in Ventral Temporal Cortex , 2001, Science.

[12]  Malte Kuss,et al.  The Geometry Of Kernel Canonical Correlation Analysis , 2003 .

[13]  Bernhard Schölkopf,et al.  Measuring Statistical Dependence with Hilbert-Schmidt Norms , 2005, ALT.

[14]  J. Berge,et al.  Tucker's congruence coefficient as a meaningful index of factor similarity. , 2006 .

[15]  Gene H. Golub,et al.  Numerical methods for computing angles between linear subspaces , 1971, Milestones in Matrix Computation.

[16]  Le Song,et al.  Supervised feature selection via dependence estimation , 2007, ICML '07.

[17]  Nikolaus Kriegeskorte,et al.  Representational Similarity Analysis – Connecting the Branches of Systems Neuroscience , 2008, Frontiers in systems neuroscience.

[18]  Keiji Tanaka,et al.  Matching Categorical Object Representations in Inferior Temporal Cortex of Man and Monkey , 2008, Neuron.

[19]  Doris Y. Tsao,et al.  Functional Compartmentalization and Viewpoint Generalization Within the Macaque Face-Processing System , 2010, Science.

[20]  W. Press.  Canonical Correlation Clarified by Singular Value Decomposition , 2011 .

[21]  StataCorp.  Stata multivariate statistics reference manual , 2011 .

[22]  J. S. Guntupalli,et al.  The Representation of Biological Classes in the Human Brain , 2012, The Journal of Neuroscience.

[23]  Mehryar Mohri,et al.  Algorithms for Learning Kernels Based on Centered Alignment , 2012, J. Mach. Learn. Res..

[24]  Kenji Fukumizu,et al.  Equivalence of distance-based and RKHS-based statistics in hypothesis testing , 2012, ArXiv.

[25]  Ha Hong,et al.  Performance-optimized hierarchical models predict neural responses in higher visual cortex , 2014, Proceedings of the National Academy of Sciences.

[26]  Surya Ganguli,et al.  Exact solutions to the nonlinear dynamics of learning in deep linear neural networks , 2013, ICLR.

[27]  Nikolaus Kriegeskorte,et al.  Deep Supervised, but Not Unsupervised, Models May Explain IT Cortical Representation , 2014, PLoS Comput. Biol..

[28]  Thomas Brox,et al.  Striving for Simplicity: The All Convolutional Net , 2014, ICLR.

[29]  Vaibhava Goel,et al.  Multimodal Retrieval With Asymmetrically Weighted Truncated-SVD Canonical Correlation Analysis , 2015, ArXiv.

[30]  Sergey Ioffe,et al.  Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift , 2015, ICML.

[31]  Yoshua Bengio,et al.  FitNets: Hints for Thin Deep Nets , 2014, ICLR.

[32]  Matthew T. Kaufman,et al.  A neural network that finds a naturalistic solution for the production of muscle activity , 2015, Nature Neuroscience.

[33]  Hod Lipson,et al.  Convergent Learning: Do different neural networks learn the same representations? , 2015, FE@NIPS.

[34]  Vaibhava Goel,et al.  Asymmetrically Weighted CCA And Hierarchical Kernel Sentence Embedding For Multimodal Retrieval , 2015 .

[35]  Nikos Komodakis,et al.  Wide Residual Networks , 2016, BMVC.

[36]  Jian Sun,et al.  Deep Residual Learning for Image Recognition , 2015, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[37]  Leon A. Gatys,et al.  Image Style Transfer Using Convolutional Neural Networks , 2016, 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).

[38]  Li Fei-Fei,et al.  Perceptual Losses for Real-Time Style Transfer and Super-Resolution , 2016, ECCV.

[39]  John P. Cunningham,et al.  Reorganization between preparatory and movement population responses in motor cortex , 2016, Nature Communications.

[40]  Samuel L. Smith,et al.  Offline bilingual word vectors, orthogonal transformations and the inverted softmax , 2017, ICLR.

[41]  Samy Bengio,et al.  Density estimation using Real NVP , 2016, ICLR.

[42]  Lukasz Kaiser,et al.  Attention is All you Need , 2017, NIPS.

[43]  Jonathon Shlens,et al.  A Learned Representation For Artistic Style , 2016, ICLR.

[44]  Jascha Sohl-Dickstein,et al.  SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability , 2017, NIPS.

[45]  Liwei Wang,et al.  Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representation , 2018, NeurIPS.

[46]  Arnold W. M. Smeulders,et al.  i-RevNet: Deep Invertible Networks , 2018, ICLR.

[47]  Xaq Pitkow,et al.  Skip Connections Eliminate Singularities , 2017, ICLR.

[48]  Masato Okada,et al.  Dynamics of Learning in MLP: Natural Gradient and Singularity Revisited , 2018, Neural Computation.

[49]  Philipp Koehn,et al.  Findings of the 2018 Conference on Machine Translation (WMT18) , 2018, WMT.

[50]  Jaehoon Lee,et al.  Deep Neural Networks as Gaussian Processes , 2017, ICLR.

[51]  Samy Bengio,et al.  Tensor2Tensor for Neural Machine Translation , 2018, AMTA.

[52]  Samy Bengio,et al.  Insights on representational similarity in neural networks with canonical correlation , 2018, NeurIPS.

[53]  Arthur Jacot,et al.  Neural tangent kernel: convergence and generalization in neural networks (invited paper) , 2018, NeurIPS.

[54]  Laurence Aitchison,et al.  Deep Convolutional Networks as shallow Gaussian Processes , 2018, ICLR.

[55]  Jaehoon Lee,et al.  Bayesian Deep Convolutional Networks with Many Channels are Gaussian Processes , 2018, ICLR.

[56]  Andrew M. Saxe,et al.  High-dimensional dynamics of generalization error in neural networks , 2017, Neural Networks.