Autoencoders for Latent Variable Modeling - Dylan Molenaar (University of Amsterdam)

Data: 08/03/2024 às 14h

Local: ICMC-USP - Sala: 3-102

Título: Autoencoders for Latent Variable Modeling

Palestrante: Dylan Molenaar - University of Amsterdam

Resumo: Latent variable models have many practical uses in modeling cross-sectional and longitudinal differences in multi-dimensional datasets. For instance in psychometrics, latent variable models are used to model individual differences and account for measurement error in theoretical constructs like intelligence and personality, and in econometrics, latent variable model are used to dynamically model econometric measures like stock wages over time. Estimation of these latent variable models has been dominated by variations of maximum likelihood estimation, least squares estimation, and Bayesian estimation. Although for continuous data, estimation is relatively fast and easily applied to high dimensional datasets, for discrete data, estimation is arguably more challenging. Therefore, recently interest has grown in estimating latent variable models using neural networks. In this talk, we focus on specific neural network: The autoencoder. That is, we study how autoencoders from the field of deep learning (Goodfellow et al., 2016) can be used to estimate latent variable models for discrete data from psychometrics in a fast and non-parametric way. Various challenges are addressed including parameter consistency, missing data, and robustness in the setup of the encoder.