Transflow Learning: Repurposing Flow Models Without Retraining

Gambardella, Andrew; Baydin, Atılım Güneş; Torr, Philip H. S.

doi:10.48550/arxiv.1911.13270

Cited by 4 publications

(5 citation statements)

References 10 publications

(15 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There are many established transfer learning techniques in prediction tasks, such as classification and recommendation . Recently, a few studies have demonstrated transfer learning methods on generative tasks for metasurface inverse designs. ,− Here, we conducted transfer learning on normalizing flow models by fine-tuning model parameters trained from the source task. Specifically, we trained the model with the data set generated from a new emitter made from tungsten.…”

Section: Resultsmentioning

confidence: 99%

Normalizing Flows for Efficient Inverse Design of Thermophotovoltaic Emitters

Yang

Fan

et al. 2023

ACS Photonics

View full text Add to dashboard Cite

Metasurfaces have shown flexibility in realizing various functionalities via shaping the geometry on the subwavelength scale. However, with increased design complexity, the flexibility makes the traditional paradigm based on expert knowledge less effective. Due to the ability to learn knowledge from raw data, deep-learning techniques have made remarkable progress in the automatic design of highperformance nanophotonic devices. However, deep-learning-based methods require a large amount of training data, which is very expensive for metasurface design. Therefore, there is still a need for efficient inverse design methods with lower data complexity and higher performance. In this Article, we modeled the inverse design problem as a probabilistic distribution learning problem and introduce normalizing flow as a flexible model to explicitly learn the distribution. To efficiently explore prior distributions without generating extra training data, we proposed a prior reshaping method that reweighs the original training data according to their fitness to the design target. The algorithm is implemented for designing niobium-based emitters in a GaSb thermophotovoltaic (TPV) cell, achieving an in-band emittance efficiency close to 99%. To further improve the data efficiency of inverse design, we explored the transfer learning ability of the proposed normalizing flow model. A new tungstenbased emitter for a GaInAsSb TPV cell was obtained with comparable performance using only 5% of data compared to the original design problem. The demonstrated approach is not only applicable to TPV device designs, but may also be used in various datadriven inverse design problems, such as artificially structured materials, synthetic materials, and mechanical devices.

show abstract

Section: Resultsmentioning

confidence: 99%

Normalizing Flows for Efficient Inverse Design of Thermophotovoltaic Emitters

Yang

Fan

et al. 2023

ACS Photonics

View full text Add to dashboard Cite

show abstract

“…The implementation of transflow-learning, which was suggested in [50], would allow the use of a trained normalizing flow on different, but similar problems without retraining the network. Such problems arise in HEP when new-physics effects from high energy scales modify scattering properties at low energies slightly and are described in an effective field theory framework.…”

Section: Discussionmentioning

confidence: 99%

i-flow: High-dimensional Integration and Sampling with Normalizing Flows

Gao,

Isaacson,

Krause

2020

Preprint

View full text Add to dashboard Cite

In many fields of science, high-dimensional integration is required. Numerical methods have been developed to evaluate these complex integrals. We introduce the code i-flow, a python package that performs high-dimensional numerical integration utilizing normalizing flows. Normalizing flows are machine-learned, bijective mappings between two distributions. i-flow can also be used to sample random points according to complicated distributions in high dimensions. We compare i-flow to other algorithms for high-dimensional numerical integration and show that i-flow outperforms them for high dimensional correlated integrals.

show abstract

“…Moreover, the multi-scale structure appears intrinsically in the distribution of natural images. Apart from generating whole images, the multi-scale representations can be used to perform other tasks, such as style transfer [47,48], content mixing [49][50][51], and texture synthesis [52][53][54][55][56].…”

Section: Related Workmentioning

confidence: 99%

RG-Flow: a hierarchical and explainable flow model based on renormalization group and sparse prior

You

et al. 2022

Mach. Learn.: Sci. Technol.

View full text Add to dashboard Cite

Flow-based generative models have become an important class of unsupervised learning approaches. In this work, we incorporate the key ideas of renormalization group (RG) and sparse prior distribution to design a hierarchical flow-based generative model, RG-Flow, which can separate information at different scales of images and extract disentangled representations at each scale. We demonstrate our method on synthetic multi-scale image datasets and the CelebA dataset, showing that the disentangled representations enable semantic manipulation and style mixing of the images at different scales. To visualize the latent representations, we introduce receptive fields for flow-based models and show that the receptive fields of RG-Flow are similar to those of convolutional neural networks. In addition, we replace the widely adopted isotropic Gaussian prior distribution by the sparse Laplacian distribution to further enhance the disentanglement of representations. From a theoretical perspective, our proposed method has $O(\log L)$ complexity for inpainting of an image with edge length $L$, compared to previous generative models with $O(L^2)$ complexity.

show abstract

Transflow Learning: Repurposing Flow Models Without Retraining

Cited by 4 publications

References 10 publications

Normalizing Flows for Efficient Inverse Design of Thermophotovoltaic Emitters

Normalizing Flows for Efficient Inverse Design of Thermophotovoltaic Emitters

i-flow: High-dimensional Integration and Sampling with Normalizing Flows

RG-Flow: a hierarchical and explainable flow model based on renormalization group and sparse prior

Contact Info

Product

Resources

About