Multimodal hierarchical Variational AutoEncoders with Factor Analysis latent space

Abstract

Real-world databases are complex, they usually present redundancy and shared correlations between heterogeneous and multiple representations of the same data. Thus, exploiting and disentangling shared information between views is critical. For this purpose, recent studies often fuse all views into a shared nonlinear complex latent space but they lose the interpretability. To overcome this limitation, here we propose a novel method to combine multiple Variational AutoEncoders (VAE) architectures with a Factor Analysis latent space (FA-VAE).Concretely, we use a VAE to learn a private representation of each heterogeneous view in a continuous latent space. Then, we model the shared latent space by projecting every private variable to a low-dimensional latent space using a linear projection matrix. Thus, we create an interpretable hierarchical dependency between private and shared information. This way, the novel model is able to simultaneously: (i) learn from multiple heterogeneous views, (ii) obtain an interpretable hierarchical shared space, and, (iii) perform transfer learning between generative models.

Automatic antibiotic resistance prediction in Klebsiella pneumoniae based on MALDI-TOF mass spectra

“…The technical results presented in this chapter are under review at Information Sciences journal from Elsevier, whose preprint can be publicly accessed [123]. In accordance with the open science philosophy upheld in this thesis, the implementation of the FA-VAE model and all associated experiments detailed in this chapter are readily accessible through a public repository on GitHub, under the link 1 .…”

“…Our second technical contribution is presented in this chapter. First, we introduce the theoretical foundations and mathematical formulation of our Factor Analysis Variational AutoEncoder (FA-VAE) [123]. Then, the following sections present experimental results to demonstrate the efficacy of the approach, including conditioning a pretrained VAE to a specific label, domain adaptation between distinct datasets and styles, and using the approach as a transfer learning tool between generative models.…”

“…In Chapter 3, a kernel-based extension, KSSHIBA [122], is proposed to handle high-dimensional data, such as that obtained from MALDI-TOF MS, while exploiting non-linear relationships in the data. Then, in Chapter 4, a VAE-based extension, FA-VAE [123], is proposed to exploit non-linear relationships and further expand the types of heterogeneous data to handle.…”

“…Therefore, in this chapter we propose a solution using ML techniques following the philosophy established in Chapter 5. To do so, we apply and tailor KSSHIBA [122] (Chapter 3) and FA-VAE [123] (Chapter 4) to the automatic ribotyping of C. diff based on their MALDI-TOF MS. We follow the same approach as in Chapter 5 using KSSHIBA with RBF, Linear and PIKE kernels. Then, we compare their performance in time, ribotyping accuracy, and latent space dimensionality to FA-VAE.…”

“…Hence, as stated before, there is a lack of reproducibility in the current state-of-the-art. In light of this gap in the literature, we propose performing automatic ribotyping of C. diff using the technical advancements presented in this dissertation, such as KSSHIBA [122] (Chapter 3) and FA-VAE [123] (Chapter 4).…”