RECORD
Generative Model with Dynamic Linear Flow
Machine Learning · Submitted 8 May 2019
abstract compiled · 2026-09-23 · no language-model review
The authors' main claim We demonstrate that the proposed DLF yields state-of-theart performance on ImageNet 32x32 and 64x64 out of all flow-based methods, and is competitive with the best autoregressive model.
The compiled record
The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
Open the compiled record
The abstract, sentence by sentence
Flow-based generative models are a family of exact log-likelihood models with tractable sampling and latent-variable inference, hence conceptually attractive for modeling complex distributions. However, flow-based models are limited by density estimation performance issues as compared to state-of-the-art autoregressive models. Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. In this paper, we propose Dynamic Linear Flow (DLF), a new family of invertible transformations with partially autoregressive structure. Our method benefits from the efficient computation of flow-based methods and high density estimation performance of autoregressive methods. We demonstrate that the proposed DLF yields state-of-theart performance on ImageNet 32x32 and 64x64 out of all flow-based methods, and is competitive with the best autoregressive model. Additionally, our model converges 10 times faster than Glow (Kingma and Dhariwal, 2018). The code is available at https://github.com/naturomics/DLF.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Flow-based generative models are a family of exact log-likelihood models with tractable sampling and latent-variable inference, hence conceptually attractive for modeling complex distributions. | background 1.00 | 0.04 | 0.03 | asserted 1.00 |
| s1 | However, flow-based models are limited by density estimation performance issues as compared to state-of-the-art autoregressive models. | background 0.73 | 0.07 | 0.02 | asserted 1.00 |
| s2 | Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. | background 0.92 | 0.05 | 0.02 | asserted 1.00 |
| s3 | In this paper, we propose Dynamic Linear Flow (DLF), a new family of invertible transformations with partially autoregressive structure. | contribution 0.91 | 0.85 | 0.12 | asserted 0.99 |
| s4 | Our method benefits from the efficient computation of flow-based methods and high density estimation performance of autoregressive methods. | contribution 0.74 | 0.32 | 0.17 | asserted 1.00 |
| s5 | We demonstrate that the proposed DLF yields state-of-theart performance on ImageNet 32x32 and 64x64 out of all flow-based methods, and is competitive with the best autoregressive model. | result 1.00 | 0.96 | 0.76 | asserted 1.00 |
| s6 | Additionally, our model converges 10 times faster than Glow (Kingma and Dhariwal, 2018). | result 1.00 | 0.94 | 0.08 | asserted 1.00 |
| s7 | The code is available at https://github.com/naturomics/DLF. | contribution 0.64 | 0.35 | 0.02 | asserted 0.98 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
The paper
Original abstract
Flow-based generative models are a family of exact log-likelihood models with tractable sampling and latent-variable inference, hence conceptually attractive for modeling complex distributions. However, flow-based models are limited by density estimation performance issues as compared to state-of-the-art autoregressive models. Autoregressive models, which also belong to the family of likelihood-based methods, however suffer from limited parallelizability. In this paper, we propose Dynamic Linear Flow (DLF), a new family of invertible transformations with partially autoregressive structure. Our method benefits from the efficient computation of flow-based methods and high density estimation performance of autoregressive methods. We demonstrate that the proposed DLF yields state-of-theart performance on ImageNet 32x32 and 64x64 out of all flow-based methods, and is competitive with the best autoregressive model. Additionally, our model converges 10 times faster than Glow (Kingma and Dhariwal, 2018). The code is available at https://github.com/naturomics/DLF.
Human review
How the literature uses it
What this paper does with the arXiv papers it cites
What it cites · 6 sentences without a confident label
Uses data or values from 2 papers
-
In particular, we use the $32×32$ and $64×64$ downsampled version of ImageNet [ref] .
Pixel Recurrent Neural Networks · Density Estimation · 0.87
-
We present samples randomly generated from the trained DLF model on ImageNet 64$×$64 and CelebA HQ 256$×$256 [ref] in Fig. [ref], both on 8-bit.
Progressive Growing of GANs for Improved Quality, Stability, and Variation · Samples and Interpolation · 0.81
Compares with 1 paper
-
To our knowledge, it is more than 10 times more efficient than Glow and Flow++ [ref] , which generally require at least thousands of epochs to converge.
Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design · Density Estimation · 1.00
Cites as background 9 papers
-
In most flow-based models [ref] , the high-dimensional random variable $$ with complex and unknown true distribution $∼p^⋆()$ is generally modeled by a latent variable $$: $= f_θ()$, where $f$ can be any bijective function with parameters $θ$ and is typically composed of a series of transformations $f=f_1∘f_2∘⋯∘f_L$. $p_θ()$ has a tractable density, such as a standard Gaussian distribution.
NICE: Non-linear Independent Components Estimation · Flow-based Models · 0.98
-
In most flow-based models [ref] , the high-dimensional random variable $$ with complex and unknown true distribution $∼p^⋆()$ is generally modeled by a latent variable $$: $= f_θ()$, where $f$ can be any bijective function with parameters $θ$ and is typically composed of a series of transformations $f=f_1∘f_2∘⋯∘f_L$. $p_θ()$ has a tractable density, such as a standard Gaussian distribution.
Density estimation using Real NVP · Flow-based Models · 0.98
-
Likelihood-based generative methods could be further divided into three different categories: variational autoencoders [ref] , autoregressive models [ref] , and flow-based generative methods [ref] .
Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling · Introduction · 0.99
Discussion
Know something about this paper? Leave a comment, or highlight any passage above to comment on it.