Traffic4cast 2020 Special Session
Friday, December 11
Session 1: 9:00-13:00 CET (0:00-4:00 PST), Session 2: 17:00-21:00 CET (8:00-12:00 PST)
Virtual-only event.
This session features an in-depth discussion of the Traffic4cast 2020 core competition results. The goal of the competition is to predict traffic in multiple big cities of different culture and economy based on industrial-scale real-world traffic data.
Traffic4cast is selected again for the prestigious Neural Information Processing Systems (NeurIPS), the leading event in machine learning. The summary of the competition is presented at the NeurIPS competition track on Friday, December 11.
Related resources:
Program
Title | Presenter | Affiliation | Duration | Abstract |
|---|---|---|---|---|
1st session: 9:00-13:00 CET (0:00-4:00 PST) session chair: Michael Kopp |
| |||
| An Introduction to the Traffic4cast Challenge | Sepp Hochreiter | IARAI, Johannes Kepler University | 15 min | |
Keynote talk: The End of Traffic and the Future of Access | David Levinson | University of Sydney | 60 min | |
| The Traffic4cast Competition Design and Data | Moritz Neun | IARAI, HERE Technologies | 30 min | |
| 1st prize: Utilizing UNet for the Future Traffic Map Prediction | Sungbin Choi | - | 40 min | |
| 2nd prize: TLab: Traffic Map Movie Forecasting Based on HR-NET | Fanyou Wu | Purdue University | 40 min | |
| 3rd prize: Towards Good Practices of U-Net for Traffic Forecasting | Jingwei Xu | Shanghai Jiao Tong University | 40 min | |
2nd session: 17:00-21:00 CET (8:00-12:00 PST) session chair: David Kreil | ||||
Keynote talk: Neural networks struggle to transfer | Razvan Pascanu | DeepMind | 60 min | |
| Graph Ensemble Net and the Importance of Feature & Loss Function Design for Traffic Prediction | Qi Qi | - | 40 min | |
| Temporal Autoencoder for Frame Prediction | Jay Santokhi | Alchera Data Technologies Ltd | 40 min | |
| Traffic Flow Prediction Using Deep Sedanion Networks | Alabi Bojesomo | Khalifa University, Abu Dhabi, UAE | 40 min | |
| Uncertainty Intervals for Graph-based Spatio-Temporal Traffic Prediction | Tijs Maas | University of Amsterdam | 40 min | |
Summary and Outlook | Michael Kopp | IARAI | 30 min |
Session Information
Keynote talk:
Prof. David P. Levinson, University of Sydney, Australia
David Levinson is a Professor of Transport Engineering at the School of Civil Engineering at the University of Sydney. He is an honorary affiliate of the Institute of Transport and Logistics Studies, where he is also a member of the Board of Advice. He received his Ph.D. in Transportation Engineering at the University of California at Berkeley in 1998. His research interests span transport, from engineering and design, through policy and planning, to geography and economics. His most recent research emphasizes transport-land use interactions, accessibility, and transport system evolution.
Abstract
Traffic—as most people have come to know it—is ending. Transport systems are being augmented with a range of information technologies. The talk discusses large scale trends that are revolutionizing the transport landscape: electrification, automation, the sharing economy, logistics, and big data. Based on all of this, I offer strategies to shape the future of infrastructure needs and priorities.
Keynote talk:
Dr. Razvan Pascanu, Deep Mind, UK
Razvan Pascanu is a research scientist at DeepMind. He received his Ph.D. from the University of Montreal in 2014. His research mainly focuses on deep learning and deep reinforcement learning. His research interests include theory of deep neural networks, graph neural networks, memory and recurrent neural networks, and learning with multiple tasks from continual learning to transfer learning, multi-task, curriculum or meta-learning.
Abstract
While we have made huge strides in improving the performance of neural networks in a static data distribution, given that we have sufficient data, there are still a lot of unanswered questions of how we can maximize transfer, and how to deal with non-stationarity. In this talk I will take a few steps back and try to describe some fundamental limitations that arise from the deep learning paradigm. In particular I will discuss how credit assignment is typically being done in neural networks (what weight gets blamed for what error), and how this mechanism, while providing amazing results, can be problematic when we are interested in transfer and building on previously acquired knowledge. I will discuss some existing directions, including concepts behind meta-learning and continual learning that could be leveraged to deal with this problem, and how maybe one can describe or formalize what we are hoping to achieve in transfer learning. While I will not be able to provide an answer to this question, I will hopefully provide new perspectives on the question of how to deal with changes in data distributions.
Traffic4cast 2020 Special Session is a free event and NeurIPS 2020 registration or account are not required to attend it.