=Paper=
{{Paper
|id=Vol-2692/xpreface
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2692/front_matter.pdf
|volume=Vol-2692
}}
==None==
https://ceur-ws.org/Vol-2692/front_matter.pdf
ADGN20: First workshop on Applied Deep
Generative Networks
Jer Hayes*, Cathal Gurrin•, Azzurra Pinio , and Mark Keane
*jeremiah.hayes@accenture.com, •cgurrin@computing.dcu.ie,
o
azzurra.pini@fjordnet.com and mark.keane@ucd.ie
https://sites.google.com/view/adgn-20/
Abstract. Generative models are widely used in many subfields of AI
and Machine Learning. More recently generative models using deep learn-
ing have been employed in a creative manner to generate new media (im-
ages, text and music), but they have also been applied to areas such as
drug discovery and data anonymisation. This workshop focused on the
applications and research problems related to the practical use of deep
generative models in the real world and where examples ranged from
drug discovery to bespoke furniture generation.
Keywords: Generative Networks · GANs · Applied Research
1 Background
A generative model is a powerful way of learning data distributions using unsu-
pervised learning; essentially the model learns unseen patterns or hidden struc-
tures in data in order to synthesis new example data. When we speak of deep gen-
erative networks we are describing models that are built as deep neural network
(or deep learning) models and that typically are grouped into types of autoen-
coders (especially variational autoencoders [1]); generative adversarial networks
(GANs) [2]; and combinations of the two. The main applications of generative
networks have been based on using image data. Given an image dataset new ex-
ample images are created, e.g., [3]. Depending on the dataset these can have
specific domains, e.g., generating realistic human faces [4]; anime characters
[5]. Another use case is ‘translation’ of one input into another. These include
transforming satellite photographs to Google maps images, transforming sketch
images to realist color photographs. A significant number of image-to-image
translations are called ‘style transfers’ and they include examples such trans-
forming photos from day to night; transforming black and white photographs to
color; transforming a painting to a photograph; and transforming a photograph
to a painting (with a certain artistic style). However, there have also been exam-
ples where text is used in association with images, e.g., text-to-image translation
where an image is created based on the text input. Some of these techniques have
also had negative applications, e.g., the use of ’deepfakes’ [6] whereby realistic
fake content is generated sometimes for purpose of creating fake news content.
2 J.Hayes et al.
The intention of the workshop was to focus on research problems related to
the practical use of deep generative models in the real world. It was also intended
to be an in-person workshop but travel restrictions due to the Covid-19 pandemic
resulted the workshop and indeed the whole conference being held virtually.
Ultimately seven papers were accepted for presentation at the workshop. These
papers explore topics mainly related the following:
– Business / Service Applications
– Exploration of latent space
– Human-machine collaboration
The papers were divided into sessions that covered business and health Ap-
plications; human-machine collaboration; and the latent Space. We now briefly
introduce these seven contributions.
2 Work presented at the workshop
Relating to business / health Applications, Pandey et al., apply GANs to the
general problem of dealing with data that is imbalanced. This occurs in many
sectors but it is especially prevalent in the financial sector where legitimate
transactions make up the vast majority of total transactions, thereby making it
harder for to detect fraudulent transactions. They presented a study on GANs
for synthetic fraud data generation and demonstrate improved classifier perfor-
mance for detecting fraud. In related work, Zhang and Li’s paper (they could not
attend the virtual workshop) introduced research on generating latent weights
for few shot image classification. This work focused on another common problem
building models from limited labeled data. Additionally Gaàl et al. presented re-
search on adding an adversarial critic model to a CNN in the application area
of Chest X-ray Lung Segmentation where both performance and interpretability
of the model results are important.
Within the domain of human + machine collaboration, Pini presented work
that takes a design approach to style transfer. This work is an application of sea-
son transfer with GANs to design a visual booking service for the travel industry.
González and Muiños-Landin Santiago then presented work on generative design
for social manufacturing with a use case featuring furniture design. Given the
remote nature of the workshop we unfortunately could not hand the physical
outputs of their use case.
Research into exploration of latent space included Singh et al. who presented
work on manifold traversal of latent spaces for novel molecule discovery. Fer-
nandes et al. also presented work on latent Space exploration for classifier im-
provement. They proposed a framework that combined GANS and evolutionary
computation to perform data augmentation on small datasets in order to improve
the performance of image classifiers trained via supervised learning.
ADGN20: First workshop on Applied Deep Generative Networks 3
3 Conclusion
Despite the vagaries of remote meetings the workshop was held as originally
intended and provided a form for both presentation and discussion. Six of the
the seven accepted papers were presented during the workshop. In terms of the
discussions that took place the focus was on the move towards real-world applica-
tion and the implications for ethics, collaboration, and interpretability of bring-
ing GANs from the lab into the real-world. It was highlighted that considering
technology application was opening up new research opportunities and that this
move to application-driven research was mirroring the situation in other fields of
AI and machine learning. Collaboration was a talking point as currently many
of these models and techniques are used by experts but Pini pointed out that
products like RunwayML (https://runwayml.com/) allow non-ML experts to use
generative networks. The question of interpretability also arose as currently for
some models it is not clear as to why a particular example was synthesised etc.
Overall the discussion was robust and involved the organisers and many of the
participants. The organisers noted the potential for follow-on workshops that
will facilitate a continuation of the discussions.
References
1. D. P. Kingma and M. Welling, “An introduction to variational autoencoders,” Foun-
dations and Trends in Machine Learning, vol. 12, no. 4, p. 307–392, 2019.
2. I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair,
A. Courville, and Y. Bengio, “Generative adversarial networks,” 2014.
3. A. Radford, L. Metz, and S. Chintala, “Unsupervised representation learning with
deep convolutional generative adversarial networks,” 2015.
4. T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for
improved quality, stability, and variation,” 2017.
5. Y. Jin, J. Zhang, M. Li, Y. Tian, H. Zhu, and Z. Fang, “Towards the automatic
anime characters creation with generative adversarial networks,” 2017.
6. R. Tolosana, R. Vera-Rodriguez, J. Fierrez, A. Morales, and J. Ortega-Garcia,
“Deepfakes and beyond: A survey of face manipulation and fake detection,” 2020.