Name: Generative Vision in AI
Start: 2018-12-10T19:00:00+01:00
End: 2018-12-10T21:00:00+01:00
Location: Brighter AI HQ

Generating images and videos with AI is one of the most exciting research fields. It drives multiple industry solutions that weren’t possible before. In this series of meetups “Generative Vision”, Brighter AI and partners will share tech knowledge, successes, failures, and opportunities to use generative neural networks. Researchers, data scientists, and technical enthusiasts are invited to join. Two company’s Machine Learning Research Leads will shed light on the latest industry research in their respective companies, Dr. Elias Vansteenkiste (Brighter AI) and Dr. Urs Bergmann (Zalando).

Meet-up Agenda:
19:00 Doors will open - Pizza and beer (sponsored by NVIDIA)
19:30 Intro, How we use AI and presenting the Meet-up theme
19:55 Q&A
20:00 Speaker - Urs Bergmann - Generative Image Models for Fashion
20:25 Q&A
20:30 Elias Vansteenkiste - Towards High Fidelity Face Generation
20:55 Q&A
21:00 Networking

Abstract topics:

Generative Image Models for Fashion - Visual Appearance is key to Fashion, with applications that are customer- and business-facing. This is why we're investigating generative image models at Zalando Research - both in an academic as well as an applied setting. I'll start my overview of our recent work by introducing our approach to generate photo-realistic textures with Generative Adversarial Networks, which allows for smooth interpolation on a manifold of textures. For use-cases in content production, but also customers wanting to try out articles, we developed generative image models that can be used to substitute articles on images of people. To make image models most useful for our fashion designers, we developed advanced feature disentanglement techniques for GANs, which allow for steering of specific attributes (color, length, etc.) or to be designed articles, and allows real-time visualization of the generated article images. Finally, our latest work on merging non-parametric modeling with parametric deep networks shows promising results in the small data regime.

Towards High Fidelity Face Generation - At Brighter, we transform visual data layer by layer to help the human or machine better understand data or even to protect the human's privacy.
We call these layers 'Perceptual Layers'. I will start with an overview of which perceptual layers we already worked on. Subsequently, I will explain how structural constraints can help to train a video transformation model unsupervised and let it produce consistent results.
Subsequently, we will dive deeper into our plan to produce high fidelity face generation.
I will start with explaining the two main components of our face model, the shape and the texture.
Next, we will move onto how we can generate the shape and the texture with conditional generative adversarial networks and on what our models should be conditioned on. In the final part, I will talk about how we combine the face model components and employ it into real-life applications.

brighter AI

Generate Vision in AI

Technology

Artificial Intelligence

Python

Machine Learning

Data Visualization

Predictive Analytics

Neural Networks

Data Science

Deep Learning

Data Science for Business

Generative Vision in AI

Brighter AI HQ

Share

Generate Vision in AI

Generative Vision in AI

Generate Vision in AI

Details

Members are also interested in