Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 242 - Deep Learning Methods in Biomedical Studies
Type: Invited
Date/Time: Tuesday, August 4, 2020 : 1:00 PM to 2:50 PM
Sponsor: WNAR
Abstract #308029
Title: Multi-Stage Sequential Deep Autoencoder-Based Monotone Nonlinear Dimensionality Reduction Methods
Author(s): Youyi Fong* and Jun Xu
Companies: Fred Hutchinson Cancer Research Center and None
Keywords: machine learning; neural network; nonlinear combination

Dimensionality reduction is a common unsupervised learning task. In this paper we propose novel deep autoencoder-based nonlinear dimensionality reduction methods by extending Kramer (1991). Our contributions are two-fold: (1) We introduce Deep Autoencoder-based Monotone (DAM) methods for promoting monotonicity between the reconstructed output and the latent components. While DAM methods do not completely solve the problem of model selection for each nonlinear component, they do reduce the model space considerably when the monotonicity assumption is reasonable. (2) We propose a new, multi-stage simultaneous (MSS) deep learning model for estimating multiple nonlinear components. This allows construction of loss curves to assess how the reconstruction error in the original input space decreases as new components are added. Using DAM and MSS together allows us to construct a loss curve that provides a more solid basis for choosing the number of nonlinear components. Our method is motivated by immune response datasets generated from vaccine studies, and will be evaluated in such datasets.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2020 program