부산대학교 도서관

The curse of dimensionality is a classic problem in machine learning which states that the number of data points required to achieve a desirable accuracy increases as the square of the dimensionality of the data points. Many dimension reduction techniques have been developed to combat this. Maximum Variance Unfolding (MVU) is one such state-of-the-art nonlinear dimension reduction algorithm that recovers a lower dimensional manifold which most of the data lies near. The algorithm's inability to handle large data sets or those that fail to meet certain distribution criteria are severe detriments to its efficiency and robustness. In this report we have introduced a variant of MVU which utilizes a novel Feed Forward Neural Network to recover the lower dimensional manifold representation. We demonstrate that our method succeeds in recovering lower dimensional manifolds and outperforms MVU in several ways including increased efficiency and flexibility.