International Workshop on
Pattern Forecasting
January 11th 2021, in conjunction with the
International Conference on Pattern Recognition (ICPR'20)
Program
The program consists of invited talks, delivered by world-famed experts in forecasting, oral presentations by the authors of accepted papers, invited spotlight presentations, brief highlights into recent possibly-influential forecasting work, and a panel discussion, where open problems in forecasting across the various fields of science would be discussed.
Schedule
The workshop starts at 12:00 CET and would conclude with a panel discussion by 16:00 CET.
Detailed schedule:
12:00 (5') Opening by workshop chairs
12:05 (20') Invited talk: Prof. Thomas Brox + Dr. Osama Makansi, University of Freiburg, DE
12:25 (20') Invited talk: Prof. Dino Zardi, University of Trento, IT
12:45 (10') Oral presentation: Wonjik Kim, Masayuki Tanaka, Masatoshi Okutomi, Yoko Sasaki. "Adaptive Future Frame Prediction with Ensemble Network"
12:55 (20') Invited talk: Prof. Carolina García Martos, Universidad Politécnica de Madrid, ES
13:15 (20') Invited talk: Prof. Giovanni Maria Farinella, University of Catania, IT
13:35 (15') Invited spotlight presentations:
Alessia Bertugli, Simone Calderara, Pasquale Coscia, Lamberto Ballan, Rita Cucchiara. "AC-VRNN: Attentive Conditional-VRNN for Multi-Future Trajectory Prediction"
Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, Nick Rhinehart. "Inverting the Pose Forecasting Pipeline with SPF2: Sequential Pointcloud Forecasting for Sequential Pose Forecasting"
Ye Yuan, Kris Kitani. "DLow: Diversifying Latent Flows for Diverse Human Motion Prediction"
Francesco Giuliari, Irtiza Hasan, Marco Cristani, Fabio Galasso. "Transformer Networks for Trajectory Forecasting"
Geri Skenderi, Marco Cristani. "Short and long-term clothing sales forecasting with exogenous factors"
Christian Joppi, Marco Cristani, Andrea Giachetti. "Finding a needle in a haystack: forecasting user search for fine-grained image retrieval"
Junwei Liang, Lu Jiang, Kevin Murphy, Ting Yu, Alexander Hauptmann. "The Garden of Forking Paths: Towards Multi-Future Trajectory Prediction"
13:50 (20') Invited talk: Prof. Marco Bee, University of Trento, IT
14:10 (20') Invited talk: Prof. Novella Bartolini, Sapienza University, IT
14:30 (10') Oral presentation: Yasuno Takato, Akira Ishii. "Rain Code: Multi-Frame Based Forecasting Spatiotemporal Precipitation Using ConvLSTM"
14:40 (20') Invited talk: Dr. Pratik Prabhanjan Brahma, Volkswagen, Belmont, CA
15:00 (20') Invited talk: Dr. James Harrison, Stanford University, CA
15:20 (40') Panel Discussion and concluding remarks
Invited Talks
Invited talks will focus on methodologies and applications of forecasting from various fields of science, to represent the topics of interest of the workshop. Invited speakers are:
University of Freiburg, DE
Title: Probabilistic Future Prediction
Abstract: Future prediction is a fundamental principle of intelligence that helps plan actions and avoid possible dangers. As the future is uncertain to a large extent, modeling the uncertainty and multimodality of the future states is of great relevance. In this talk, I will discuss our recent works addressing the aforementioned challenges and their applications (especially in autonomous driving). Moreover, the talk will conclude by highlighting some important findings about the distributions of common real datasets and potential future directions.
University of Trento, IT
Title: Identification of atmospheric circulation patterns for climate and weather predictions
Abstract: Climate and weather are inherently difficult to predict for a series of reasons. First of all, despite the impressive progress achieved by observational technologies, our capability of determining precisely the relevant variables (e. g. wind speed and direction, air temperature, water vapor concentration, etc.) at all the relevant scales is still quite limited. Even our capability of predicting these variables by means of numerical models suffers from many limitations, arising from both uncertainties and inaccuracies in the initial and boundary conditions, and truncation errors, and, mostly, in the intrinsic growth of errors associated with the strongly nonlinear character of the equations governing atmospheric processes. However there are many aspect of weather and climate variability than can be captured in terms of “modes”, as the status of the mean values of many atmospheric variables associated with these modes is quite determined in a narrow range of variability. This is the case for example of circulation patterns associated with typical configurations of the atmospheric pressure field, viewed in terms of the layout of isobars at sea level pressure, or of level contours of isobaric surfaces. Further phenomena, such as rainfall or cloud cover, are closely associated with each of these patterns. An analysis of recurrency and strength of these patterns allows higher predictability skills, and appreciable indirect forecast of the associated phenomena.
Universidad Politécnica de Madrid, ES
Title: A review on Time Series Forecasting using Dynamic Factor Analysis: applications in Energy and Economics
Abstract: In this talk the use of dimensionality reduction techniques for forecasting high-dimensional vector of series is presented. The starting point will be the papers by Peña and Box (1987) and Lee and Carter (1992). Then, I will present some extensions of these works, particularly a summary of several papers I have co-authored in the last 10 years. In most of them the application is related to energy markets, particularly electricity prices forecasting. But there are other real data examples related to CO2 emissions and fossil fuel prices as well as wind power production or Industrial Production Indexes (IPI).
University of Catania, IT
Title: Towards Future Predictions in Egocentric Perception
Abstract: The ability to predict the future is fundamental for humans to explore the world, learn, perceive, navigate, and act. It is hence expected that future AI systems shall be able to reproduce such abilities, reasoning about the world in terms of events and stimuli which live in the near or distant future. Predictive abilities are also fundamental for wearable systems to understand the user’s short- and long-term goals, offer appropriate guidance based on the user’s objectives, and improve user’s safety anticipating future actions. In this talk, I will present recent research on future predictions from egocentric video which has been carried out at the Image Processing Laboratory (IPLAB) at the University of Catania, Italy. We will first introduce the main motivations behind research on egocentric perception, then discuss approaches to predict future interacted objects and actions from egocentric video. The talk will also focus on the relevant datasets to support the study of future prediction tasks from egocentric video, such as the EPIC-KITCHENS series of datasets and challenges, the TREK-100 dataset which allow the study of visual object tracking in First Person Vision, as well as our newly introduced MECCANO dataset for studying human-object interaction recognition and future prediction in industrial-like scenarios. References: http://iplab.dmi.unict.it/fpv/
University of Trento, IT
Title: Realized Peaks over Threshold: a Time-Varying Extreme Value Approach with High-Frequency based Measures
Abstract: Recent contributions to the financial econometrics literature exploit high-frequency (HF) data to improve models for daily asset returns. We propose a new class of dynamic extreme value models that profit from HF data when estimating the tails of daily asset returns. Our realized peaks-over-threshold approach provides estimates for the tails of the time-varying conditional return distribution.
Sapienza University, IT
Title: AI for finance: from trading strategies to market manipulation
Abstract: Algorithmic trading accounts for more than 70% of equity trade volume in the US market. Algorithmic traders execute their strategies at ultra high speed contributing to the ever-increasing market efficiency, and leaving little room for short-term profit to human traders. Designing predictive models to extract signals for trading opportunities in a world dominated by algorithmic traders is therefore a challenging task. In addition, the automated use of predictive techniques exposes traders to market manipulation. Interdependency among financial assets is one of the relevant aspects to factor in the design of predictive tools. It is in fact observed that price shocks propagate across interdependent markets. We will discuss reasons for shock propagation and show tools for predicting them, discussing related opportunities, risks and potential unfairness among investors.
Volkswagen, Belmont, CA
Title: Challenges in Building Scalable Prediction Modules for Autonomous Vehicles
Abstract: Predicting behavior and future trajectory patterns of surrounding vehicles, pedestrians and other road users is extremely crucial for safely operating autonomous vehicles. In this talk, I will enumerate some of the unique aspects of the prediction problem that makes it very different from other components like perception and planning in a typical autonomous driving modular stack. Given its probabilistic and inter-dependent nature in an multi-agent setting, we will go through some of the challenges in reliably training and subsequently evaluating such multi-output models using commonly used loss functions and metrics. I will also talk about some of our recent efforts in doing simultaneous perception and intention prediction of all pedestrians in the scene in a single shot manner.
Stanford University, CA
Title: Connecting Meta-Learning and Dynamic Models through Differentiable Filtering
Abstract: Meta-learning has seen substantial recent success in few-shot and data-constrained learning. By backpropagating through an online learning algorithm, meta-learning aims to learn an initialization and/or update rule such that performance given a small amount of data is maximized. However, the standard meta-learning formulation has a number of drawbacks; chiefly, the setting is focused on few-shot episodic learning in a static environment and is interested primarily in point estimates. These assumptions stand in stark contrast to time series forecasting, in which predictions are made with continuously changing unobserved factors, and characterization of uncertainty is critical. In this talk, we present a collection of simple meta-learning methods based on Bayesian linear regression and Gaussian discriminant analysis. These methods induce minimal additional complexity in the learning process, and are capable of rapid adaptation and uncertainty quantification during online learning. Next, we generalize these methods to dynamic models via a simple class of differentiable filtering algorithms, which show strong performance in time-varying learning problems. Finally, we discuss extensions of these methods to problems with discrete changepoint behavior, leveraging differentiable changepoint detection. We demonstrate our methods on a wide variety of forecasting and time-varying supervised learning problems