Measuring HSIC(Hilbert Space Independence Critation) between intermediate layers in Deep Neural Network models.
The experimental content is to measure the HSICs of each layer in the pre-trained MNIST classifier (DNN model). The model classifies the MNIST test data (10,000 images-class samples) and the state of each layer is taken.
input | lay1 | act1 | lay2 | act2 | lay3 | act3 | lay4 | act4 | lay5 | act5 | |
---|---|---|---|---|---|---|---|---|---|---|---|
input | 2.95E-06 | 0.0001690372894 | 1.45E-05 | 0.000141198873 | 0.0000469996998 | 0.0001351601098 | 0.00008241862139 | 0.0001126127641 | 0.00002569158032 | 0.000009351987592 | 0.0000000004710106661 |
lay1 | 0.000263440744 | 0.01245696657 | 0.001357019575 | 0.01041677749 | 0.004523675079 | 0.009755719057 | 0.007451573117 | 0.00940600974 | 0.002617222478 | 0.0009929085876 | 0.00000005291088778 |
act1 | 0.00003277336582 | 0.001357019575 | 0.0002527607864 | 0.001726789366 | 0.0008697763892 | 0.001633875156 | 0.001444036774 | 0.001718668348 | 0.0004998143853 | 0.0001834386443 | 0.000000009532376531 |
lay2 | |||||||||||
Act2 | |||||||||||
lay3 | |||||||||||
act3 | |||||||||||
lay4 | |||||||||||
act4 | |||||||||||
lay5 | |||||||||||
act5 | 0.000001583722253 | 0.00000003030216742 | 0.000001212205375 | 0.00000003181505095 | 0.0000008012884461 | 0.00000006116697143 | 0.000002186409585 | 0.00000145169179 | 0.000006354812953 | 0.000006504984405 | 0.000000002807068279 |
-
"Dimensionality Reduction for Supervised Learning with Reproducing Kernel Hilbert Spaces", (JMLR 2004)
https://www.di.ens.fr/~fbach/fukumizu04a.pdf -
"Measuring Statistical Dependence with Hilbert-Schmidt Norms", (ALT 2005)
http://www.gatsby.ucl.ac.uk/~gretton/papers/GreBouSmoSch05.pdf -
"A Kernel Statistical Test of Independence", (NIPS 2007)
https://papers.nips.cc/paper/3201-a-kernel-statistical-test-of-independence.pdf -
"Kernel Methods for Deep Learning", (NIPS 2009)
https://papers.nips.cc/paper/3628-kernel-methods-for-deep-learning.pdf -
"Deep Gaussian Process", (AISTATS 2013)
http://proceedings.mlr.press/v31/damianou13a.pdf -
"Deep Neural Networks as Gaussian Process", (ICLR 2018)
https://arxiv.org/pdf/1711.00165.pdf -
"Stronger generalization bounds for deep nets via a compression approach", (ICML 2018)
http://proceedings.mlr.press/v80/arora18b/arora18b.pdf