DSPT#56 - How to forecast at scale and pull out the big GANs (Aveiro)

Details

We are going from time-series forecast at scale, showing some scenarios of going from small to larger clients problems, to GAN's, were you can do useful representation learning or use it to fake things.

This meetup is sponsored by Bosch (https://www.bosch.pt/).

=== SCHEDULE ===

The preliminary agenda for the meetup is the following:
• 18:30-19:00: Welcome and get together
• 19:00-19:30: Time Series forecast at scale - André Antunes
• 19:40-19:45: Group photo
• 19:45-20:15: Networking / Coffee Break - by Ubiwhere
• 20:15-20:45: Pulling out the big GANs: from useful representation learning to faking things - Eduardo Pinho
• 20:50: Closing
• 21:00: Dinner is optional but it might be an excellent opportunity for networking. (register here: https://doodle.com/poll/nkyxhgi5idbykxh6 )

Talk 1
Title: Time Series forecast at scale

Abstract: Going from one small customer to 2 of the 3 biggest customers in the Portuguese market came as a great challenge. For the amount of work we came across, we needed to make sure that the same model creation procedure could be applied to different scenarios, in different geographical locations, for different demographics, faster and with at least the same performance as before. In this presentation I’m going to guide you on how we managed to scale up our development process and how you can do it as well.

Short bio: André Antunes is MSc in Mechanical Engineering from University of Aveiro (2017). He has been working at SCUBIC as lead data scientist since then, focusing mainly on machine learning for time series forecast, hydraulic simulation and optimization.

Talk 2:
Title: Pulling out the big GANs: from useful representation learning to faking things

Abstract: The modern practice of deep learning has enabled automatic solutions to problems once thought very hard to achieve. Its emergence in multiple computer science fields (computer vision, natural language processing, and more) is influencing a wide spectrum of projects and applications. In a recent breakthrough, generative adversarial networks (GANs) made way for use cases beyond surprisingly appealing generative models, and they could become the culprit for many of the upcoming attempts to fool our society and create fake information. Yet, much of their success can be traced back to the study of representation learning as a superset of deep learning.

This presentation will first cover a brief overview of representation learning, followed by a few noteworthy methods. The second part provides an explanation of what GANs are and what they can do, up to the current state of the art. Before the conclusion, a use case for unsupervised learning methods is presented in the context of information extraction from medical images.

Short bio: Eduardo Pinho is a finalizing PhD student in computer science at IEETA, with a focus on multimodal information retrieval systems in medical imaging repositories. Research interests include medical imaging informatics, content-based image retrieval, and deep learning methods applied to medical images. He is also a maintainer of multiple OSS projects, including Dicoogle, an open-source medical imaging archive used worldwide.
Twitter: @e_net4
GitHub: github.com/Enet4