α-EM algorithm and α-ICA learning based upon extended logarithmic information measures

The α-logarithm extends the logarithm as the special case of α = -1. Usage of α-related information measures based upon this extended logarithm is expected to be effective to speedup of convergence, i.e., on the improvement of learning aptitude. In this paper, two typical cases are investigated. One is the α-EM algorithm (α-Expectation-Maximization algorithm) which is derived from the α-log-likelihood ratio. The other is the α-ICA (α-Independent Component Analysis) which is formulated as minimizing the α-mutual information. In the derivation of both algorithms, the α-divergence plays the main role. For the α-EM algorithm, the reason for the speedup is explained using Hessian and Jacobian matrices for learning. For the α-ICA learning, methods of exploiting the past and future information are presented. Examples are shown on single-loop α-EM's and sample-based α-ICA's. In all cases, effective speedups are observed. Thus, this paper's examples together with formerly reported ones are evidences that the speed improvement by the α-logarithm is a general property beyond individual problems.