Wednesday, September 21, 3.15 pm, 2022

Learning Neural Causal Models
Stefan Bauer, Division of Decision and Control Systems, KTH Royal Institute of Technology
Abstract: Many questions in everyday life as well as in research are causal in nature: How would the climate change if we lower train prices or will my headache go away if I take an aspirin? Inherently, such questions need to specify the causal variables relevant to the question and their interactions. However, existing algorithms for learning causal graphs from data are often not scaling well, both with the number of variables or the number of observations. This talk will provide a brief introduction to causal structure learning, recent efforts in using continuous optimization to learn causal graphs at scale and systematic approaches to generate interventional data to identify the data generating mechanism as quickly as possible.
Location: Ada Lovelace
Organizers: Fredrik Lindsten, Sourabh Balgi

Wednesday, October 12, 3.15 pm, 2022

Title: Human Motion and Dynamics Capture with Limited Data
Bastian Wandt, Department of Electrical Engineering (ISY), Linköping University
Abstract: 3D human motion capture has a wide range of applications across society with the most common examples being in movie production, medicine, autonomous driving, and human-robot interaction. In recent years, rapid deep learning driven advances in research have translated directly into commercial products. However, despite successful commercialization, many problems remain for these data-hungry machine learning approaches. Arguably the most important problem is the lack of training data covering the whole variety of human motions; it is even questionable whether such data can ever be recorded. This talk will present several solutions for monocular human motion capture that side-step the problem of missing data by the use of weakly- or unsupervised machine learning techniques in combination with geometric modeling. It will also give a glimpse on future applications in motion analysis by capturing human dynamics, i.e. forces and torques on and inside the human body.
Location: Ada Lovelace
Organizer: Fredrik Lindsten, Sourabh Balgi

Wednesday, November 9, 3.15 pm, 2022

Title: Learning-based acceleration of scientific computations
Jens Sjölund, Department of Information Technology, Uppsala University
Abstract: In this talk I will ask not what scientific computations can do for machine learning, but what machine learning can do for scientific computations. Computational models and simulators - many of them exceptionally demanding - are ubiquitous in science and engineering. I will introduce the general idea behind learning-based acceleration of scientific computations and survey recent works focusing on partial differential equations and optimization problems. After highlighting some of the limitations of these works I will, finally, argue that geometric deep learning is the way forward.
Location: Ada Lovelace
Organizer: Fredrik Lindsten, Sourabh Balgi

Wednesday, December 7, 3.15 pm, 2022

Title: The Beneficial Role of Stochastic Noise in SGD
Hans Kersting, SIERRA Team, INRIA and Ecole Normale Superieure, Paris
Abstract: The data sets used to train modern machine-learning models are often huge, e.g., millions of images. This makes it too expensive to compute the true gradient over all data sets. In each gradient descent (GD) step, a stochastic gradient is thus computed over a subset ("mini-batch") of data. The resulting stochastic gradient descent (SGD) algorithm, and its variants, is the main workhorse of modern machine learning. Until recently, most machine-learning researchers would have preferred to use GD, if they could, and considered SGD only as a fast approximation to GD. But new research suggests that the stochasticity in SGD is part of the reason why SGD works so well. In this talk, we investigate multiple theories on the advantages of the noise in SGD, including better generalization in flatter minima ('implicit bias') and faster escapes from difficult parts of the landscapes (such as saddle points and local minima). We highlight how correlating noise can help optimization and zoom in on the question which noise structure would be optimal for SGD.
Related reading: Anticorrelated Noise Injection for Improved Generalization , Explicit Regularization in Overparametrized Models via Noise Injection
Location: Zoom and Ada Lovelace
Organizers: Fredrik Lindsten, Sourabh Balgi