Theoretical Error Performance Analysis for Deep Neural Network Based Regression Functional Approximation

Thumbnail Image
Qi, Jun
Lee, Chin-Hui
Ma, Xiaoli
Associated Organizations
Supplementary to
Based on Kolmogorov's superposition theorem and universal approximation theorems by Cybenko and Barron, any vector-to-scalar function can be approximated by a multi-layer perceptron (MLP) within certain bounds. The theorems inspire us to exploit deep neural networks (DNN) based vector-to-vector regression. This dissertation aims at establishing theoretical foundations on DNN based vector-to-vector functional approximation, and bridging the gap between DNN based applications and their theoretical understanding in terms of representation and generalization powers. Concerning the representation power, we develop the classical universal approximation theorems and put forth a new upper bound to vector-to-vector regression. More specifically, we first derive upper bounds on the artificial neural network (ANN), and then we generalize the concepts to DNN based architectures. Our theorems suggest that a broader width of the top hidden layer and a deep model structure bring a more expressive power of DNN based vector-to-vector regression, which is illustrated with speech enhancement experiments. As for the generalization power of DNN based vector-to-vector regression, we employ a well-known error decomposition technique, which factorizes an expected loss into the sum of an approximation error, an estimation error, and an optimization error. Since the approximation error is associated with our attained upper bound upon the expressive power, we concentrate our research on deriving the upper bound for the estimation error and optimization error based on statistical learning theory and non-convex optimization. Moreover, we demonstrate that mean absolute error (MAE) satisfies the property of Lipschitz continuity and exhibits better performance than mean squared error (MSE). The speech enhancement experiments with DNN models are utilized to corroborate our aforementioned theorems. Finally, since an over-parameterized setting for DNN is expected to ensure our theoretical upper bounds on the generalization power, we put forth a novel deep tensor learning framework, namely tensor-train deep neural network (TT-DNN), to deal with an explosive DNN model size and realize effective deep regression with much smaller model complexity. Our experiments of speech enhancement demonstrate that a TT-DNN can maintain or even achieve higher performance accuracy but with much fewer model parameters than an even over-parameterized DNN.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI