Japjot Singh

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-162

May 20, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-162.pdf

In the past decade, Machine Learning research has grown tremendously. The increased availability of data and powerful hardware has brought forward many applications in different industries. Generative modeling, specifically synthetic data generation, has made headlines with models capable of creating fake celebrity images and deep fakes. Normalizing Flows are one family of generative models with desirable qualities, including exact density estimation and inexpensive sampling. Unlike other generative modeling techniques like generative adversarial networks, variational autoencoders, and autoregressive models, Normalizing Flows show impressive results on both image and non-structured tabular data indicating their effectiveness in modeling complex distributions. Although still in relative infancy, they have shown promising results when used with other models or as a synthetic data source for separate downstream tasks.

This thesis explores applications in computer vision-based detection of COVID-19 and supervisory planning in reinforcement learning for energy demand response. Our work in the healthcare application presents a hybrid conditional generative model which decouples feature representations from input images to generate quality artificial samples in label scarce domains, which prove to be effective in several downstream tasks. We further investigate the flexibility of normalizing flow methods to capture energy price responses within a proprietary reinforcement learning environment and use this in a hybrid planning model scheme which in turn improves the learning and performance of a price controlling agent.

Advisors: Costas J. Spanos

---

BibTeX citation:

@mastersthesis{Singh:EECS-2022-162,
    Author= {Singh, Japjot},
    Title= {Generative Modeling for Healthcare Applications and Energy Demand Response with Normalizing Flows},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-162.html},
    Number= {UCB/EECS-2022-162},
    Abstract= {In the past decade, Machine Learning research has grown tremendously. The increased availability of data and powerful hardware has brought forward many applications in different industries. Generative modeling, specifically synthetic data generation, has made headlines with models capable of creating fake celebrity images and deep fakes. Normalizing Flows are one family of generative models with desirable qualities, including exact density estimation and inexpensive sampling. Unlike other generative modeling techniques like generative adversarial networks, variational autoencoders, and autoregressive models, Normalizing Flows show impressive results on both image and non-structured tabular data indicating their effectiveness in modeling complex distributions. Although still in relative infancy, they have shown promising results when used with other models or as a synthetic data source for separate downstream tasks. 

This thesis explores applications in computer vision-based detection of COVID-19 and supervisory planning in reinforcement learning for energy demand response. Our work in the healthcare application presents a hybrid conditional generative model which decouples feature representations from input images to generate quality artificial samples in label scarce domains, which prove to be effective in several downstream tasks. We further investigate the flexibility of normalizing flow methods to capture energy price responses within a proprietary reinforcement learning environment and use this in a hybrid planning model scheme which in turn improves the learning and performance of a price controlling agent.},
}

EndNote citation:

%0 Thesis
%A Singh, Japjot 
%T Generative Modeling for Healthcare Applications and Energy Demand Response with Normalizing Flows
%I EECS Department, University of California, Berkeley
%D 2022
%8 May 20
%@ UCB/EECS-2022-162
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-162.html
%F Singh:EECS-2022-162