Transcript Deep Learning - University of Houston
Kernel Analysis of Deep Networks
By: Gregoire Montavon Mikio L. Braun Klaus-Robert Muller (Technical University of Berlin)
JMLR 2011
Presented by: Behrang mehrparvar (University of Houston)
April 8th, 2014
Roadmap
Deep Learning Goodness of Representations Measuring goodness Role of architecture
Deep Learning?
Distributed representation Depth Less examples in regions Capture global structure Efficient representation Abstraction Flexibility Higher-level features Incorporate prior knowledge
Distributed Representation [1]
Depth [2]
Abstraction [?]
Problem Specification
Deep Learning is still a
Black Box!
Theoretical aspect e.g. studying depth in sum-product networks Analytical arguments e.g. analysis of depth Experimental results e.g. performance in application domains Visualization e.g. measuring invariance
Kernel Methods
Decouples learning algorithms from data representation Kernel operator: Measures similarity between points All the prior knowledge of the learning problem In this paper: Not a learning machine Abstraction tool to model the deep network
Kernel Methods (cont.)
Kernel Methods model the deep network Used to quantify ...
the goodness of representations the evolution of good representations
Hypothesis
1) Simpler and more accurate representation throughout the depth 2) Structure of the network (restrictions) define the speed of how representations are formed – Evolution from dist. of pixels to dist. of classes
Problem Specification
Problem
: Role of depth in goodness of representation
Challenge
: Definition and Measurement for goodness
Solution
: – Simplicity – Dimensionality: number of kernel PCs Number of local variations Accuracy Classification error
Hypothesis (Cont.)
Method
1) Train the deep network 2) Infer the representation of each layer 3) Apply kernel PCA on each layer representations 4) Project data points on first d eigenvectors 5) Analyze the results
Method (Analysis)
Why Kernels?
1) Incorporating prior knowledge 2) Measurable simplicity and accuracy 3) Theoretical framework and convergence bounds [3] 4) Flexibility
Dimensionality and Complexity
Dimensionality and Complexity (cont.)
Intuition
Accuracy – Task-relevant information Simplicity – Number of allowed local variations in the inputs space – However, does not explain domain-specific regularities – Robust to number of samples • vs. number of support vectors
Effects of Kernel mapping
Experiment setup
Datasets – – Tasks MNIST CIFAR – – Supervised learning Transfer learning Architectures – – – Multilayer perceptron (MLP) Pretrained multilayer perceptron (PMLP) Convolutional neural networks (CNN)
Effect of Settings
Effect of Depth (Hyp. 1)
Observation
Higher layers – – More accurate representations More simple representations
Architectures
Multilayer Perceptrons – – No preconditioning on learning problem
Prior: NONE
Pretrained Multilayer perceptrons – – – Better represents the underlying representation Contains a certain part of soluton
Prior: generative model of input
Convolutional Neural Networks –
Prior: Spatial invariance
Multilayer Perceptron [4]
Convolutional Neural Networks [4]
Effect of Architecture (Hyp. 2)
Observation
MNIST: – – MLP: Discriminating is solved greedily PMLP and CNN: postpone to last layers CIFAR – – MLP: Doesn't discriminate till last layer PMLP and CNN: spread it to more layers
WHY?!
– –
Good observation, but no explanation!
Hints: dataset, priors, etc. ?
Effect of Architecture (Cont.)
Observation
Regularities in PMLP and CNN – Facilitate the construction of a structured solution – Controls the rate of discrimination at every level
Label Contribution of PCs
Comments
Strengths – – – – Important and interesting Simple and intuitive Well designed Good
analysis approach experiments
of results
problem
Weaknesses – – Too many observations • e.g. role of sigma in scale invariance explaining observations
Future works?
Experiments on
Unsupervised Learning
Explaining
the results Analysis on biological neural systems?!
References
1) Bengio, Yoshua, and Olivier Delalleau. "
On the expressive power of deep architectures.
" Algorithmic Learning Theory. Springer Berlin Heidelberg, 2011.
2) Poon, Hoifung, and Pedro Domingos. "
Sum-product networks: A new deep architecture.
" Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on. IEEE, 2011.
3) Braun, Mikio L., Joachim M. Buhmann, and Klaus Robert Müller. "
On relevant dimensions in kernel feature spaces.
" The Journal of Machine Learning Research 9 (2008): 1875-1908.
4) http://deeplearning.net/
Thanks ...