Search
mlsplogo MLSP2016
IEEE International Workshop on
Machine Learning for Signal Processing

September 13-16, 2016  Vietri sul Mare, Salerno, Italy

Towards Socially-aware AI: Human Behavior Understanding at Scale
Alexandre Alahi Postdoc Alexandre Alahi
Stanford Vision Lab & Computational Vision and Geometry Lab,
Stanford University, USA
Personal homepage

Biography:

Alexandre Alahi is currently a postdoctoral fellow at Stanford University and received his PhD from EPFL. His research interests span visual information processing, computer vision, machine learning, robotics, and are focused around understanding and forecasting human social behaviors. In particular, he is interested in sparse approximation, deep learning, big visual data processing, real-time machine vision, and multi-modal reasoning. He was awarded the Swiss NSF early and advanced researcher grants. He won the CVPR 2012 Open Source Award for his work on Retina-inspired image descriptor, and the ICDSC 2009 Challenge Prize for his sparsity driven algorithm to track sport players. His PhD was nominated for the EPFL PhD prize. His research has been covered by the Wall Street journal, and PBS/abc TV channel in the US, as well as European newspapers, Swiss TV news, and Euronews TV channel in Europe. He has also co-founded the startup Visiosafe, transferring his research on understanding human behavior into an industrial product. He was selected as the Top 20 Swiss Venture leaders in 2010 and won several startup competitions.

Abstract:

Over the past sixty years, Intelligent Machines have made great progress in playing games, tagging images in isolation, and recently making decisions for self-driving vehicles. Despite these advancements, they are still far from making decisions in social scenes and assisting humans in public spaces such as terminals, malls, campuses, or any crowded urban environment. To overcome these limitations, we need to empower machines with social intelligence, i.e., the ability to get along well with others and facilitate mutual cooperation. This is crucial to design smart spaces that adapt to the behavior of humans for efficiency, or develop autonomous machines that assist in crowded public spaces (e.g., delivery robots, or self-navigating segways).

In this talk, I will present my work towards socially-aware machines that can understand human social dynamics and learn to forecast them. First, I will highlight the machine vision techniques behind understanding the behavior of more than 100 million individuals captured by multi-modal cameras in urban spaces. I will show how to use sparsity promoting priors to extract meaningful information about human behavior from an overwhelming volume of high dimensional and high entropy data. Second, I will introduce a new deep learning method to forecast human social behavior. The causality behind human behavior is an interplay between both observable and non-observable cues (e.g., intentions). For instance, when humans walk into crowded urban environments such as a busy train terminal, they obey a large number of (unwritten) common sense rules and comply with social conventions. They typically avoid crossing groups and keep a personal distance to their surrounding. I will present detailed insights on how to learn these interactions from millions of trajectories. I will describe a new recurrent neural network that can jointly reason on correlated sequences and simulate human trajectories in crowded scenes. It opens new avenues of research in learning the causalities behind the world we observe. I will conclude my talk by mentioning some ongoing work in applying these techniques to social robots, and the first generation of smart hospitals.


Gaussian Processes for Signal Processing
Richard E. Turner Lecturer Richard E. Turner
Department of Engineering
University of Cambridge, United Kingdom
Personal homepage

Biography:

Richard Turner holds a Lectureship (equivalent to US Assistant Professor) in Computer Vision and Machine Learning in the Computational and Biological Learning Lab, Department of Engineering, University of Cambridge, UK. Before taking up this position, he held an EPSRC Postdoctoral research fellowship which he spent at both the University of Cambridge and the Laboratory for Computational Vision, NYU, USA. He has a PhD degree in Computational Neuroscience and Machine Learning from the Gatsby Computational Neuroscience Unit, UCL, UK and a M.Sci. degree in Natural Sciences (specialism Physics) from the University of Cambridge, UK. His research interests include machine learning for signal processing and developing probabilistic models of perception.

Abstract:

Gaussian Processes are receiving growing attention in the machine learning community as they provide flexible methods for placing distributions over functions and performing inference and learning directly in this function space. Gaussian Processes formally equivalent to neural networks with a single infinite hidden layer and often outperform their finite-dimensional cousins. This talk will provide an intuitive introduction to Gaussian Processes before outlining a number of emerging research frontiers in this area that lie on the interface with signal processing.

I will begin by describing how the humble Gaussian distribution can be used for non-linear regression before formalising this approach in terms of a Gaussian Process. I will then connect Gaussian Processes to classical signal processing methods including the short-time Fourier transform (STFT), spectrogram, filter bank, and wavelet representations of signals. I will show how the new perspective allows standard machine learning tools to be leveraged, transforming these classical approaches into uncertainty-aware, adaptive, and hierarchical representations that can better restore signals and synthesise realistic synthetic versions, notably audio textures such as howling wind, falling rain, and running water. Next, I will show how Gaussian Processes can be used for non-parametric spectral estimation. Finally, I will talk about recent work on Deep Gaussian Processes, which are equivalent to multi-layer neural networks with infinite numbers of neurons in each hidden layer, that have achieved state-of-the-art results on a large number of regression tasks.


Toward Causal Machine Learning
Bernhard Schölkopf Director Bernhard Schölkopf
Department of Empirical Inference
Max Planck Institute for Intelligent Systems
Tübingen, Germany
Home page

Biography:

Dr. Schölkopf received his doctorate in computer science from the Technical University Berlin (1997) and is heading the Department of Empirical Inference at Max Planck Institute for Intelligent Systems, Tübingen.

His scientific interests are in the field of machine learning and inference from empirical data. In particular, the study kernel methods for extracting regularities from possibly high-dimensional data. These regularities are usually statistical ones, however, in recent years his has also become interested in methods for finding causal structures that underly statistical dependences. He has worked on a number of different applications of machine learning - in data analysis, you get "to play in everyone's backyard." Most recently, he has been trying to play in the backyard of astronomers and photographers.

Abstract:

In machine learning, we use data to automatically find dependences in the world, with the goal of predicting future observations. Most machine learning methods build on statistics, but one can also try to go beyond this, assaying causal structures underlying statistical dependences. Can such causal knowledge help prediction in machine learning tasks? We argue that this is indeed the case, due to the fact that causal models are more robust to changes that occur in real world datasets. We touch upon the implications of causal models for machine learning tasks such as domain adaptation, transfer learning, and semi-supervised learning. We also present an application to the removal of systematic errors for the purpose of exoplanet detection. Machine learning currently mainly focuses on relatively well-studied statistical methods. Some of the causal problems are conceptually harder, however, the causal point of view can provide additional insights that have substantial potential for data analysis.

Powered by CONWIZ, © Copyright 2015 | Privacy Policy and Terms of Use | Page maintained by Jan Larsen