Abstract
Over the past few years, there has been increasing attention on the self-supervised learning of depth and motion estimation from a sequence of monocular images. Despite the ill-posed property of estimating depth or scene flow from monocular images, designed proxy losses enable to learn to estimate them without dense annotation. However, the accuracy of the current approaches highly depends on the proxy losses, which becomes one of the main limitations. In this talk, we are going to review the literature on estimating depth, optical flow and scene flow using monocular images, their proxy loss designs with limitations, and possible future direction.

Bio
Junhwa Hur has been working as a Ph.D. student at Technische Universität Darmstadt, advised by Prof. Stefan Roth. He received his MS degree from Seoul National University in 2013. His main research interests lie on dynamic scene understanding (e.g., jointly estimating motion, depth, and semantic segmentation) using self-supervised learning. He has been acknowledged as one of the outstanding reviewers multiple times in CVPR or ECCV.

Abstract
In this talk, we propose an analysis of gradient descent on wide two-layer ReLU neural networks that leads to sharp characterizations of the learned predictor. The main idea is to study the dynamics when the width of the hidden layer goes to infinity, which is a Wasserstein gradient flow. While this dynamics evolves on a non-convex landscape, we show that its limit is a global minimizer if initialized properly. We also study the "implicit bias" of this algorithm when the objective is the unregularized logistic loss. We finally discuss what these results tell us about the generalization performance. This is based on joint work with Francis Bach.

Bio
Lénaïc Chizat is a CNRS researcher at Université Paris-Saclay. He obtained his PhD in applied mathematics at Université Paris-Dauphine in 2017. He is working on the mathematical analysis of data-driven algorithms. His current research interests include the theory of optimal transport and the theory of artificial neural networks.

Abstract
We study the problem of phase retrieval in which one aims to recover a function from the magnitude of its wavelet transform. It is already known that if the wavelet is a Cauchy wavelet, then the modulus of the wavelet transform uniquely determines the analytic part of any square-integrable function. We consider bandlimited functions and derive new uniqueness results for phase retrieval, where the wavelet itself can be complex-valued. In particular, we prove the first uniqueness result for the case that the wavelet has a finite number of vanishing moments. In addition, we establish the first result on unique reconstruction from samples of the wavelet transform magnitude when the wavelet coefficients are complex-valued.

Bio
Francesca Bartolucci is a Postdoc at the Department of Mathematics at ETH Zürich. She received her PhD at the University of Genova under the supervision of Filippo De Mari and Ernesto De Vito with the thesis entitled "Radon transforms: Unitarization, Inversion and Wavefront sets". Her research interests include phase retrieval problems, group representations theory, wavelet analysis, shearlet analysis, microlocal analysis and Radon transforms.

Abstract
For robots to become an effective component of our society, it is necessary that these agents become primarily cognitive systems, endowed with a cognitive architecture that enables them to adapt, acquire experience, act and interact pro-actively and predictively with the environment and communicate with the human partners. Human communication depends on mutual understanding: I know how to communicate because I entertain and adapt a model of you, which enables me to select an effective way to convey to you what I want and to have an intuition of your internal states – what you need, prefer, fear or desire. Such intuition enables me to perceive properties that would be otherwise not accessible to my perception, as goals, emotions or effort, and to establish a common ground, a shared perception with the partner. Our contribution to the roadmap toward the architecture of cognitive systems leverages on the use of a humanoid robot (iCub) as a tool of investigation of human cognitive abilities. Moreover, the robot plays a crucial role to test some of our assumptions on how to build a cognitive interactive agent. We attempt at modeling the minimal skills necessary for cognitive development, starting from the visual features that enable to recognize the presence and the behavior of other agents in the scene and to foster automatic coordination in human-robot interactive tasks. In a dual approach, we are trying to understand how to modulate robot behavior to elicit better human understanding and to express different characteristics of the interaction: from the mood to the level of commitment. This approach is propaedeutic to the creation of a cognitive system, by helping in the definition of what is relevant to attend to, starting from signals originating from the intrinsic characteristics of the human body. In parallel, we are investigating novel learning frameworks, to enable the robot to adapt autonomously to the partner’s needs and preferences, to foster the establishment of a solid relation with the machine, going beyond the single interaction. We believe that only a structured effort toward cognition will in the future lead to more humane machines, able to see the world and people as we do and engage with them in a meaningful manner.

Bio
Alessandra received her Ph.D. in Humanoid Technologies from the University of Genova (Italy) in 2010. After a Post Doc at the Italian Institute of Technology (IIT) and two research periods in USA and Japan, she became the scientific responsible of the Cognitive Robotics and Interaction Laboratory of the RBCS Dept. at IIT. After being Assistant Professor in Bioengineering at DIBRIS University of Genoa, she is now Tenure-Track Researcher at the Italian Institute of Technology, head of the COgNiTive Architecture for Collaborative Technologies (CONTACT) unit. In 2018 she has been awarded the ERC Starting Grant wHiSPER, focused on the investigation of joint perception between humans and robots. She published more than 60 papers and abstracts and participated in the coordination of the CODEFROR European IRSES project. She is an Associate Editor of Robots and Autonomous Systems, Cognitive Systems Research and the International Journal of Humanoid Robotics and she has served as a member of the Program Committee for the International Conference on Human-Agent Interaction and IEEE International conference on Development and Learning and Epigenetic Robotics. The scientific aim of her research is to investigate the sensory and motor mechanisms underlying mutual understanding in human-human and human-robot interaction.

Abstract
Sparsity properties for phase-space representations of several types of operators (including pseudodifferential, metaplectic and Fourier integral operators) have been extensively studied in recent articles, with applications to the analysis of dispersive evolution equation. It has been proved that such operators are approximately diagonalized by Gabor wave packets - equivalently, the corresponding phase-space representations (Gabor matrix/kernel) can be thought of as sparse infinite-dimensional matrices. While wave packets are expected to undergo some spreading and dispersion phenomena, there is no record of these issues in the aforementioned estimates. We recently proved refined estimates for the Gabor matrix of metaplectic operators, also of generalized type, where sparsity, spreading and dispersive properties are all simultaneously noticeable. We also provide applications to the propagation of singularities for the Schrödinger equation; in this connection, our results can be regarded as a microlocal refinement of known estimates.

Bio
I am a Ph.D. candidate in Mathematics from the joint program by Politecnico and University of Turin. I joined the CHARML Unit of the MaLGa Center, University of Genoa, as a Postdoctoral fellow starting from October 2020. Broadly speaking, my research interests focus on modern harmonic analysis and its applications to mathematical physics and PDEs.

Abstract
I will present a novel convolutional neural network designed for learning pseudodifferential operators in the context of linear inverse problems. Such a network is able to replicate and outperform the results of the Iterative Soft Thresholding Algorithm (ISTA), a well-known reconstruction algorithm in sparsity-promoting minimization problems. By a combination of techniques and tools from regularization theory of inverse problems, multi-resolution wavelet analysis, and the theory of pseudodifferential operators, we are able to theoretically deduce the architecture of the network and to prove its convergence properties. Our case study is limited-angle computed tomography: we test two different implementations of our network on simulated data from limited-angle geometry, achieving noteworthy preliminary results. This is a joint project with T. A. Bubba, M. Lassas, S. Siltanen from University of Helsinki and M. Galinier, M. Prato from Università di Modena.

Bio
I got my Ph.D. at Politecnico di Milano in February 2019. Since March 2019 I have been a post-doc researcher in the Inverse Problems group at the University of Helsinki. My research focuses on inverse problems related to (nonlinear) PDEs, optimization, and combining regularization theory with machine learning approaches.

Abstract
In this talk I will start questioning the fact that semantic and temporal annotations are taken for granted in action recognition. I will show that this leads to some issues and will present two works to tackle such problems. I will then briefly talk about EPIC Kitchens, the largest Egocentric video dataset to which I contributed during my PhD, focusing mainly on how we annotated it. Finally, I will present my last work which proposes a weakly supervised method for action recognition using a novel type of temporal annotations, i.e. single timestamps roughly aligned with the actions.

Bio
I took my PhD in Computer Science at the University of Bristol (UK) in November 2019. In March 2020 I joined the Nanyang Technological University of Singapore as a research fellow. My area of research is action recognition and video understanding. In particular my interest focuses on weakly supervised approaches and the link between labels and learning.

Abstract
In this talk, we present simple ideas to combine nonparametric approaches based on positive definite kernels with deep learning models. There are many good reasons for bridging these two worlds. On the one hand, we want to provide regularization mechanisms and a geometric interpretation to deep learning models, as well as a functional space that allows to study their theoretical properties (e.g., invariance and stability). On the other hand, we want to bring more adaptivity and scalability to traditional kernel methods, which are crucially lacking. We will start this presentation by introducing models to represent graph data, then move to biological sequences, and images, showing that our hybrid models can achieves state-of-the-art results for many predictive tasks, especially when large amounts of annotated data are not available. This presentation is based on joint works with Alberto Bietti, Dexiong Chen, and Laurent Jacob.

Bio
Julien Mairal is a research scientist at Inria Grenoble, where he leads the Thoth research team. He joined Inria Grenoble in 2012, after a post-doc in the statistics department of UC Berkeley. He received the Ph.D. degree from Ecole Normale Superieure, Cachan. His research interests include machine learning, computer vision, mathematical optimization, and statistical image and signal processing. In 2016, he received a Starting Grant from the European Research Council. He was awarded the Cor Baayen prize in 2013, the IEEE PAMI young researcher award in 2017 and the test-of-time award at ICML 2019

Abstract
It is a well-known fact in the classical theory of PDE's that harmonic functions in Euclidean domains can be characterized in terms of the mean value property. Namely, a continuous function u is harmonic if and only if the value of u at x coincides with its average over any ball centered at x. In recent years, similar mean value characterizations have been obtained for p-harmonic functions, which are defined as weak solutions of the so-called p-Laplace equation. This interplay between equations and mean value properties has turned out to be the cornerstone in the development of new approximation techniques for the study of certain properties of the p-Laplace equation, such as the regularity of solutions. In this talk we briefly review some of these results.

Abstract
Plankton is at the bottom of the food chain. Microscopic phytoplankton account for about 50% of all photosynthesis on Earth, corresponding to 50 billion tons of carbon each year, or about 125 billion tonnes of sugar. Thus, monitoring plankton is paramount to infer potential dangerous changes to the ecosystem. In this work we describe an application of machine learning to environmental sensing which uses plankton as biosensor. An important bottleneck of the adoption of state of the art machine learning tools into cell biology is the need for large high quality annotations. We introduce a pipeline to perform accurate detection and classification of plankton species with minimal supervision. Our algorithms approach the performance of existing supervised machine learning algorithms when tested on a plankton dataset generated from a custom-built lensless digital device. We propose a set of morphological features to establish the baseline for using plankton as a biosensor. Using an anomaly detection approach, we show that it is possible to detect deviation from the average space of features for plankton microorganisms, that we propose could be related to environmental threat or perturbations. Such an approach can open the way for the development of an automatic Artificial Intelligence (AI) based system for using plankton as biosensor.

Abstract
Limited angle geometry is still a rather challenging modality in computed tomography (CT), in which entire boundary sections are not captured in the measurements making the reconstruction a severly ill-posed inverse problem. Compared to the standard filtered back-projection, iterative regularization-based methods help in removing artifacts but still cannot deliver satisfactory reconstructions. Based on the result that limited tomographic data sets reveal parts of the wavefront (WF) set in a stable way and artifacts from limited angle CT have directional properties, we present a hybrid reconstruction framework that combines model-based sparse regularization with data-driven deep learning. The core idea is to solve the compressed sensing formulation associated to the limited angle CT problem to recover the so called “visible” part of WF and learning via a convolutional neural network architecture the “invisble” ones, which provably cannot be handled by model-based methods. Such a decomposition into visible and invisible parts is achieved using the shearlet transform that allows to resolve WF sets in the phase space. Our numerical experiments show that our approach surpasses both pure model- and more data-based reconstruction methods, while offering an (heuristic) understanding of why the method works, providing a more reliable approach especially for medical applications.

Abstract
In the deep learning era, to address common 3D scene and object understanding applications we delegate a neural network the task of extracting feature representations from data and to learn intermediate subspaces and embeddings. This dangerously exposes our algorithms to known limitations of neural networks, in particular the need for huge amount of labeled data and the domain shift. In this talk, I will walk through current approaches for using deep learning for 3D reconstruction and 3D recognition tasks, and highlight novel directions to solve these limitations. In the first part of the talk, I will give an overview of methods aimed at monocular 6D object pose estimation and discuss how current techniques are tackling the problem of domain shift. In the second part of the talk, I will focus instead on 3D scene understanding and monocular 3D reconstruction, discussing the use of monocular depth prediction for monocular SLAM and related applications, as well as looking at approaches for unsupervised and self-supervised learning.

Bio
Federico Tombari is a research scientist and manager at Google and a lecturer at the Technical University of Munich (TUM). He has more than 180 refereed papers in the field of computer vision, machine learning and robotic perception. He got his PhD in 2009 from the University of Bologna, at the same institution he was Assistant Professor from 2013 to 2016. In 2008 and 2009 he was an intern and consultant at Willow Garage, California. Since 2014 he has been leading a team of PhD students at TUM on computer vision and deep learning. In 2017-2018, he was co-founder and managing director of Pointu3D Gmbh, a Munich-based startup on 3D perception for AR and robotics. He was the recipient of two Google Faculty Research Award (in 2015 and 2018) and an Amazon Research Award (in 2017). He has been a research partner, among others, of Google, Toyota, BMW, Audi, Amazon, Stanford and JHU. His works have been awarded at conferences and workshops such as 3DIMPVT'11, MICCAI'15, ECCV-R6D'16, AE-CAI'16, ISMAR '17

Abstract
During 20th century, the problem of inverting Radon transform has been deeply studied because of its various applications. A classical use in the continuous case (e.g. two-dimensional signals) is the medical tomography. Recently, Radon on discrete setups (e.g. graphs) also began to be treated and network tomography has been developed with the same philosophy of algorithms used in CT scan. We analyze Radon on a discrete manifold: a homogeneous tree X. On X the set H of all the horocycles is defined, they are the analog of hyperplanes in Euclidean spaces. Given a signal f on X, its Radon transform Rf is defined at a horocycle h as the sum of values of f at vertices lying on h. We solves the unitarization problem: we show the existence of a pseudo-differential operator such that its postcomposition with R extends to a unitary operator Q from L^2(X) to L^2(H). Furthermore, such Q is an intertwining operator for the representations of the automorphism group of X on L^2(X) and the one on L^2(H), which are therefore unitary equivalent.

Abstract
In this talk we introduce a modern approach to manifold valued prediction based on structured prediction techniques. We present a statistically consistent estimator for the task of learning functions whose output space is a differentiable manifold. We then show applications on various manifolds ranging from imaging to robotics showing the flexibility of the proposed approach.

Abstract
Recent advances on markerless pose estimation based on computer vision and deep neural networks are opening the possibility of adopting efficient methods for extracting precise human pose and movement information from video data. In this paper we report the results of a pilot study carried out on a clinical gait analysis study-case, where we compare 2D parameters computed with a reference marker-based technique with the ones obtained with a markerless pipeline. The results we report are encouraging as they show there are no statistically significant differences between a set of selected parameters computed with the standard approach and the markerless one. Our study opens to a wide range of application of the approach on the variety of clinical domains, with countless benefits in terms of simplicity, unobtrusiveness, and computational efficiency.

Abstract
Central objects of classical Fourier Analysis are the Fourier transform, convolution operators, periodic and non-periodic functions and so on. Distribution theory widens the scope by allowing larger families of Banach spaces of functions or generalized functions and extending many of the concepts to this more general setting. Although, according to A. Weil the natural setting for Fourier Analysis (leading to the spirit of Abstract Harmonic Analysis: AHA) most of the time one works in the setting of the Schwartz space of rapidly decreasing functions and its dual space, the tempered distributions. In this setting weighted Lp-spaces and Sobolev spaces correspond to each other in a very natural way. In this talk we will summarize the advantages with respect to the level of technical sophistication and theoretical background which is possible when one uses instead of the Schwartz-Bruhat space the Segal algebra S0(G) and the resulting Banach Gelfand Triple, which appears to be suitable for the description of most problems in AHA as well as for many engineering applications (this part is beyond the scope of the current talk). Among others the use of Wiener amalgam spaces and modulation spaces (introduced by the author in the 1980s) belong to a comprehensive family of Banach spaces, which we call Fourier Standard Spaces. These spaces have a double module structure, with respect to convolution by integrable functions and pointwise multiplication with functions from the Fourier algebra.

Abstract
We want to explore a different approach to the duality in the entropic optimal transport, much more in the spirit of optimal transport, which is different from the usual techniques coming from the Schrodinger problem. This will result in consistent a priori estimates, which are preserved in the limit of epsilon going to 0. As a byproduct we find new estimates for the Schrodinger problem and we prove the IPFP algorithm is converging also in the multimarginal case.

Bio
Simone Di Marino is an Associate Professor at University of Genova. He obtained his PhD at Scuola Normale Superiore in 2014 and his main research interests concern optimal transport and its applications. Previously he has held a post-doctoral position at University Paris-Sud 11 and he was a permanent Indam researcher at Scuola Normale Superiore.

Abstract
In this talk we will visit the landscape of dictionary learning via iterative thresholiding and K residual means. For a givengenerating dictionary we will have a look at the basin of attraction, the regions of contraction, and spurious attractive points. Time permitting we will also discuss heuristics how to use escape from spurious attractive points and jump directly into the basin of attraction.

Bio
Karin Schnass holds an MSc in Mathematics from the University of Vienna (AT), 2004, and a PhD in Computer, Communication and Information Sciences from EPFL (CH), 2009. Following a postdoc at RICAM Linz (AT) and two maternity leaves, she spent 2 years as Schroedinger fellow at the University of Sassari (IT). In 2015 she joined the University of Innsbruck first as assistant professer, since 2019 as associate professor, where she is heading the FWF-START-project ‘Optimisation Principles, Models & Algorithms for Dictionary Learning' (6 yrs, 1.2M€). She is an expert on (theoretical) dictionary learning.

Abstract
More and more machine learning systems make data-driven decisions in the real work, in increasingly higher-stakes applications. This has caused substantial interest in reinforcement learning -- the field of learning to make decisions from data -- which has seen stunning recent empirical breakthroughs. At its heart is the challenge of trading exploration -- collecting data for learning better models -- and exploitation -- using the estimate to make decisions. In many applications, however, exploration is a potentially dangerous proposition, as it requires experimenting with actions that have unknown consequences. Hence, most prior work has confined exploration to simulated environments. In this talk, I will present our work towards rigorously reasoning about safety of exploration in reinforcement learning. I will discuss a model-free approach, where we seek to optimize an unknown reward function subject to unknown constraints. Both reward and constraints are revealed through noisy experiments, and safety requires that no infeasible action is chosen at any point. I will also discuss model-based approaches, where we learn about system dynamics through exploration, yet need to verify safety of the estimated policy. Our approaches use Bayesian inference over the objective, constraints and dynamics, and -- under some regularity conditions -- are guaranteed to be both safe and complete, i.e., converge to a natural notion of reachable optimum. I will also show experiments on safely tuning cyber-physical systems in a data-driven manner.

Bio
Andreas Krause is a Professor of Computer Science at ETH Zürich, where he leads the Learning & Adaptive Systems Group. He also serves as Academic Co-Director of the Swiss Data Science Center. Before that he was an Assistant Professor of Computer Science at Caltech. He received his Ph.D. in Computer Science from Carnegie Mellon University (2008) and his Diplom in Computer Science and Mathematics from the Technical University of Munich, Germany (2004). He is a Microsoft Research Faculty Fellow and a Kavli Frontiers Fellow of the US National Academy of Sciences. He received ERC Starting Investigator and ERC Consolidator grants, the Deutscher Mustererkennungspreis, an NSF CAREER award, the Okawa Foundation Research Grant recognizing top young researchers in telecommunications as well as the ETH Golden Owl teaching award. His research on machine learning and adaptive systems has received awards at several premier conferences and journals. Andreas Krause served as Program Chair for ICML 2018, and is serving as Action Editor for the Journal of Machine Learning Research.

Abstract
In everyday life we are surrounded by various sensors, wearable and not, that explicitly or implicitly record information on our behavior either visible and hidden (e.g. physiological activity). Such sensors are of different nature: accelerometer, gyroscope, camera, electrodermal activity sensor, heart rate monitor, breath rate monitor and others. Most important, the multimodal nature of data is apt to sense and understand the many facets of human daily-life behavior from physical, voluntary activities to social signaling and lifestyle choices influenced by affect, personal traits, age and social context. The intelligent sensing community is able to exploit the data acquired with these sensors in order to develop machine-learning-based techniques, which can help in improving predictive models of human behavior. In this talk will be presented the latest research at Imaging and Vision Laboratory (http://www.ivl.disco.unimib.it/) in the field of human behavior monitoring, both at the sensing and the understanding levels, by using multimodal data sources. Applications of interest can relate to domotics, healthcare, transport, education, safety aid, entertainment, sports and others.

Bio
Paolo Napoletano is assistant professor of Computer Science (tenure track - RTDB) at Department of Informatics, Systems and Communication of the University of Milano-Bicocca. In 2007, he received a Doctor of Philosophy degree (PhD) in Information Engineering from the University of Salerno (Italy) with a thesis focused on Computational Vision and Pattern Recognition. In 2003, he received a Master's degree in Telecommunications Engineering from the University of Naples Federico II, with a thesis focused on Transmission of Electromagnetic Fields. His current research interests focus on signal, image and video analysis and understanding, multimedia information processing and management and machine learning for multi-modal data classification and understanding. More information at http://www.ivl.disco.unimib.it/people/paolo-napoletano/.