Robust PCA learning rules based on statistical physics approach

Abstract:  Abstract A statistical physics approach is adapted to the problem of robust principal component analysis (PCA). Some commonly used PCA learning rules are connected to some energy function, which is further generalized by adding a binary decision field with a given prior distribution so that outliers are considered. The generalized energy is used to define a Gibbs distribution and to derive an effective energy function, which is further used to derive a learning rule for robust, PCA. Experimental results have shown that the robust rules considered have improved the performance of the PCA algorithms significantly