Latent Variable Models
By Simon Washington, Matthew G. Karlaftis, Fred L. Mannering, Panagiotis Ch. Anastasopoulos. 2019
We make the following striking observation: fully convolutional VAE models trained on 32×32 ImageNet can generalize well, not just to 64×64 but also to far larger photographs, with no changes to the model. We use this property, applying fully convolutional models to lossless compression, demonstrating a method to scale the VAE-based ‘Bits-Back with ANS’ algorithm for lossless compression (Townsend et al., 2019) to large color photographs, and achieving state of the art for compression of full size ImageNet images. We release Craystack, an open source library for convenient prototyping of lossless compression using probabilistic models, along with full implementations of all of our compression results1.
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
- David C. Minnen
- Saurabh Singh
- Eirikur Agustsson
- Nick Johnston
- Johannes Ballé
- Sung Jin Hwang
- Fabian Mentzer
- S. Freyaldenhoven
Citations in the corpus
- Variational image compression with a scale hyperprior (2018, background): “A standard way to model dependencies between a set of target variables is to introduce latent variables conditioned on which the target variables are assumed to be independent (Bishop, 1999).”
- Unsupervised Machine Learning for Networking: Techniques, Applications and Research Challenges (2017, background): “Latent variable model allows us to express relatively complex distributions in terms of tractable joint distributions over an expanded variable space [95].”
- Pre-Event Trends in the Panel Event-Study Design (2019, methodology): “Replacement of η it with x it produces a factor model or measurement error model (Aigner et al. 1984). A large literature, partially reviewed in Abbring and Heckman (2007), Heckman and Vytlacil (2007), and Matzkin (2007), shows how to establish identification in such models, typically by imposing co”
- Teaching Who You Are: Connecting Teachers’ Civic Education Ideology to Instructional Strategies (2018, background): “However, many researchers have called into question the use of these cutoffs (Barrett, 2007; Kline, 2011; Loehlin, 2004;Wolf, Harrington, Clark, & Miller, 2013).”