| 概要(Abstract) |
Dimensionality reduction is demanding in high-dimensional data analysis to extract meaningful insights. Although Automatic Relevance Determination (ARD) offers a solution to automatically reduce the dimensionality in a latent variable model, i.e., Bayesian PCA, it falters when handling noisy or nonlinear datasets, such as those from calcium imaging. We introduce a dual ARD method that applies ARD priors to both loading weights and latent variables. Through evaluations on both simulated and real calcium imaging datasets, our dual ARD consistently outperformed conventional dimensionality reduction methods, especially in handling non-linear observations. Specifically, in actual calcium imaging data, the dual ARD effectively identified key low-dimensional latent variables that were adequate for decoding tasks. |