Friday, April 14, 2017

Understanding Deep Neural Networks; Opening the Black Box, Energy Propagation in DNNs, Exploring loss function topology

Today, we have three papers trying to explain why deep neural networks work.  From the first paper:
"In our experiments, most of the optimization epochs are spent on compressing the internal representations under the training error constraint. This compression occurs by the SGD without any other explicit regularization or sparsity, and - we believe - is largely responsible for the absence of overfitting in DL. This observation also suggests that there are many (exponential in the number of weights) different randomized networks with essentially optimal performance. "


Despite their great success, there is still no comprehensive theoretical understanding of learning with Deep Neural Networks (DNNs) or their inner organization. Previous work [Tishby & Zaslavsky (2015)] proposed to analyze DNNs in the Information Plane; i.e., the plane of the Mutual Information values that each layer preserves on the input and output variables. They suggested that the goal of the network is to optimize the Information Bottleneck (IB) tradeoff between compression and prediction, successively, for each layer.
In this work we follow up on this idea and demonstrate the effectiveness of the Information-Plane visualization of DNNs. We first show that the stochastic gradient descent (SGD) epochs have two distinct phases: fast empirical error minimization followed by slow representation compression, for each layer. We then argue that the DNN layers end up very close to the IB theoretical bound, and present a new theoretical argument for the computational benefit of the hidden layers.

Many practical machine learning tasks employ very deep convolutional neural networks. Such large depths pose formidable computational challenges in training and operating the network. It is therefore important to understand how many layers are actually needed to have most of the input signal's features be contained in the feature vector generated by the network. This question can be formalized by asking how quickly the energy contained in the feature maps decays across layers. In addition, it is desirable that none of the input signal's features be "lost" in the feature extraction network or, more formally, we want energy conservation in the sense of the energy contained in the feature vector being proportional to that of the corresponding input signal. This paper establishes conditions for energy conservation for a wide class of deep convolutional neural networks and characterizes corresponding feature map energy decay rates. Specifically, we consider general scattering networks, and find that under mild analyticity and high-pass conditions on the filters (which encompass, inter alia, various constructions of Weyl-Heisenberg filters, wavelets, ridgelets, (α)-curvelets, and shearlets) the feature map energy decays at least polynomially fast. For broad families of wavelets and Weyl-Heisenberg filters, the guaranteed decay rate is shown to be exponential. Our results yield handy estimates of the number of layers needed to have at least ((1−ε)⋅100)% of the input signal energy be contained in the feature vector.

Exploring loss function topology with cyclical learning rates by Leslie N. Smith, Nicholay Topin
We present observations and discussion of previously unreported phenomena discovered while training residual networks. The goal of this work is to better understand the nature of neural networks through the examination of these new empirical results. These behaviors were identified through the application of Cyclical Learning Rates (CLR) and linear network interpolation. Among these behaviors are counterintuitive increases and decreases in training loss and instances of rapid training. For example, we demonstrate how CLR can produce greater testing accuracy than traditional training despite using large learning rates. Files to replicate these results are available at this https URL

Join the CompressiveSensing subreddit or the Google+ Community or the Facebook page and post there !
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.

No comments: