Data Mapping and Finite Difference Learning
This work provides a unified framework for dimensionality reduction and feature extraction in machine learning, though it appears incremental as it builds on existing RBM concepts.
The paper tackles the challenge of training Restricted Boltzmann Machines (RBMs) by proposing a data mapping approach that minimizes squared error on the visible layer using finite difference learning, enabling real-valued matrix data, non-sigmoid activations, and showing that contrastive divergence is an approximation of gradient descent.
Restricted Boltzmann machine (RBM) is a two-layer neural network constructed as a probabilistic model and its training is to maximize a product of probabilities by the contrastive divergence (CD) scheme. In this paper a data mapping is proposed to describe the relationship between the visible and hidden layers and the training is to minimize a squared error on the visible layer by a finite difference learning. This paper presents three new properties in using the RBM: 1) nodes on the visible and hidden layers can take real-valued matrix data without a probabilistic interpretation; 2) the famous CD1 is a finite difference approximation of the gradient descent; 3) the activation can take non-sigmoid functions such as identity, relu and softsign. The data mapping provides a unified framework on the dimensionality reduction, the feature extraction and the data representation pioneered and developed by Hinton and his colleagues. As an approximation of the gradient descent, the finite difference learning is applicable to both directed and undirected graphs. Numerical experiments are performed to verify these new properties on the very low dimensionality reduction, the collinearity of timer series data and the use of flexible activations.