An Introduction to Variational Autoencoders

Kingma, Diederik P.; Welling, Max

doi:10.1561/9781680836233

Cited by 1,419 publications

(2,008 citation statements)

References 0 publications

Supporting

Mentioning

1,447

Contrasting

Unclassified

Order By: Relevance

“…where q i is the model of the mean number of coincidences in the ith line of response (LOR) (or sinogram bin). Next, it is necessary to define an objective function which indicates how well the parameters x of the model for (1) correspond to the actual measured data, modeled by (2) and (3). The goal of image reconstruction is then to find the parameter vector x, for (1), which when forward modeled with (2), best agrees with the acquired noisy measured data (3), according to a chosen objective (or cost) function as follows:…”

Section: A Basic Principlesmentioning

confidence: 99%

“…In the training datasets, for the case of supervised learning, example inputs are paired with their corresponding desired outputs. For unsupervised learning, the training data may consist of example inputs only (for learning of latent representations of the data [3]), or of unpaired example inputs and example outputs [4]. A further category, that of self-supervised learning [5], [6], needs only input data examples and instructions on how to create labels (rather than providing labels) thus reducing the need for human interaction with the learning process.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Deep Learning for PET Image Reconstruction

Reader

Corda-DrIncan

Mehranian

et al. 2021

IEEE Trans. Radiat. Plasma Med. Sci.

162

View full text Add to dashboard Cite

This article reviews the use of a subdiscipline of artificial intelligence (AI), deep learning, for the reconstruction of images in positron emission tomography (PET). Deep learning can be used either directly or as a component of conventional reconstruction, in order to reconstruct images from noisy PET data. The review starts with an overview of conventional PET image reconstruction and then covers the principles of general linear and convolution-based mappings from data to images, and proceeds to consider nonlinearities, as used in convolutional neural networks (CNNs). The direct deep-learning methodology is then reviewed in the context of PET reconstruction. Direct methods learn the imaging physics and statistics from scratch, not relying on a priori knowledge of these models of the data. In contrast, model-based or physics-informed deep-learning uses existing advances in PET image reconstruction, replacing conventional components with deep-learning data-driven alternatives, such as for the regularization. These methods use trusted models of the imaging physics and noise distribution, while relying on training data examples to learn deep mappings for regularization and resolution recovery. After reviewing the main examples of these approaches in the literature, the review finishes with a brief look ahead to future directions.

show abstract

Section: A Basic Principlesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Deep Learning for PET Image Reconstruction

Reader

Corda-DrIncan

Mehranian

et al. 2021

IEEE Trans. Radiat. Plasma Med. Sci.

162

View full text Add to dashboard Cite

show abstract

“…They were then suggested by Dimmick et al 9 as tools for training super resolution networks by using the features extracted by passing Hi-C data through a trained autoencoder as a loss function. In this manuscript we expand upon this strategy, but replace their network with a different flavor of network called the variational autoencoder 11 .…”

Section: (Eq 1)mentioning

confidence: 99%

“…While previous work often split chromosomes into training, validation and testing sets in a sequential manner 8 9 we were concerned that differences in the 3D conformation of large vs small chromosomes 16 may contain implicit bias in contact map features that could confound training. Consequently we assembled training, validation and test sets in a non sequential manner using chromosomes 1,3,5,6,7,9,11,12,13,15,17,18,19,21 as our training set, chromosome 2,8,10,22 as our validation set and chromosomes 4,14,16,20 as our test set.…”

Section: Dataset Assemblymentioning

confidence: 99%

VEHiCLE: a Variationally Encoded Hi-C Loss Enhancement algorithm

Highsmith

Cheng

2020

Preprint

View full text Add to dashboard Cite

Chromatin conformation plays an important role in a variety of genomic processes. Hi-C is one of the most popular assays for inspecting chromatin conformation. However, the utility of Hi-C contact maps is bottlenecked by resolution. Here we present VEHiCLE, a deep learning algorithm for resolution enhancement of Hi-C contact data. VEHiCLE utilises a variational autoencoder and adversarial training strategy to enhance contact maps, making them more viable for downstream analysis. VEHiCLE expands previous efforts at Hi-C super resolution by providing novel insight into the biologically meaningful and human interpretable feature extraction. Using a variational autoencoder VEHiCLE provides a user tunable, full generative model for generating synthetic Hi-C data while also providing state-of-the-art results in enhancement of Hi-C data across multiple metrics.

show abstract

“…A specific type of black-box model that is targeted to capturing the physics of a specific problem, is the generative model. More specifically, Generative Adversarial Networks (GANs) [7] and Variational Autoencoders (VAEs) [8] are neural networks that learn how to generate data that look like reality. It is believed that, by studying these types of neural networks and the way they produce data, further understanding of the underlying problems may be achieved.…”

Section: How Much Useful (Safe) Life Remains (Prognosis)?mentioning

confidence: 99%

An Application of Generative Adversarial Networks in Structural Health Monitoring

Tsialiamanis¹,

Chatzi²,

Dervilis³

et al. 2020

XI International Conference on Structural Dynamics

View full text Add to dashboard Cite

In the current work, the use of generative adversarial networks (GANs) in a simulated structural health monitoring (SHM) application is studied. A specific type of GAN is considered, aiming at a disentangled representation of underlying features and clusters of data through some latent variables. This idea could prove useful in SHM, since explanation of how damage mechanisms or environmental conditions affect a structure may be exploited in order to monitor structures more effectively. In a simulated mass-spring example, different damage cases are introduced by reducing the stiffness of specific springs and different damage levels by applying different extents of stiffness reduction. The GAN implementation proves able to capture different damage cases through its categorical latent variables, as well as the damage extent within its continuous latent variables. The results demonstrate that the latent variables are indeed capturing the effect of damage in the structure and can be exploited for the purpose of condition assessment.

show abstract

An Introduction to Variational Autoencoders

Cited by 1,419 publications

References 0 publications

Deep Learning for PET Image Reconstruction

Deep Learning for PET Image Reconstruction

VEHiCLE: a Variationally Encoded Hi-C Loss Enhancement algorithm

An Application of Generative Adversarial Networks in Structural Health Monitoring

Contact Info

Product

Resources

About