Talks by Visitors

Photo Gonzalo Muñoz

Gonzalo Muñoz (Universidad de O’Higgins, Chile) [homepage]
Coordinates: @ Zuse Institute Berlin (ZIB), Room 2006 (Seminar Room)

Title. A Sample Size-Efficient Polyhedral Encoding and Algorithm for Deep Neural Network Training Problems

Abstract. (click to expand)

Deep Learning has received much attention lately, however, results regarding the computational complexity of training deep neural networks have only recently been obtained. Moreover, all current training algorithms for DNNs with optimality guarantees possess a poor dependency on the sample size, which is typically the largest parameter. In this work, we show that the training problems of large classes of deep neural networks with various architectures admit a polyhedral representation whose size is linear in the sample size. This provides the first theoretical training algorithm with provable worst-case optimality guarantees whose running time is polynomial in the size of the sample.

Photo Ariel Liebmann

Ariel Liebmann (Monash University, Australia) [homepage]
Coordinates: @ Zuse Institute Berlin (ZIB), Room 2006 (Seminar Room)

Title. Optimisation, Machine Learning and AI for Rapid Grid Decarbonisation

Abstract. (click to expand)

The national and transcontinental electricity grids of today are based on devices such as coal furnaces, steam turbines, copper and steel wires, electric transformers, and electromechanical power switches that have remained unchanged for 100 years. However imperceptibly, the components and operational management of this great machine, the grid, has began to change irreversibly. This is fortunate, as climate science tells us we must reduce CO2 emissions from the energy sector to zero by 2050 and to 50% of current levels by 2030 if we are to prevent dangerous climate changes in future world that is over 1.5 degree hotter that today. Now utility scale wind and solar PV farms as large as coal, gas and nuclear generators are being deployed more cheaply than it is possible to build and operate generators using older technologies. In some cases, even these new technologies can be cheaper that even merely the operating costs of older technologies. In addition, low cost rooftop solar PV has also enabled consumers to become self-suppliers and also contributors to the supply of energy for their neighbours. Moreover, the “dumb” grid of the past, is becoming “smarter”. This is enabled through a combination of ubiquitous low-cost telecommunication and programmable devices at the edge of the grid such as smart meters, smart PV inverters, smart air conditioners and home energy management systems. The final component is the electrification of the private transport system that will finally eliminate the need for fossil fuels. The implications of this are that it is now necessary to rapidly replan and reinvest in the energy system at rates and in ways that are unprecedented in industrial civilisations history. While the majority of hardware technology already exist, the missing piece of the puzzle are new computers science technologies, and particularly Optimisation, Machine Learning, Forecasting and Data analytics methods needed to plan and operate this rapidly transforming system.

In this talk I will describe a range of ways existing computer science tools in the Optimisation, AI, ML and other areas we and others are enhancing in order to better operate and plan the existing power system. I will focus on identifying emerging research opportunities in areas that are needed to complete the transformation to a cheaper, smarter and zero carbon energy system.

Photo Masashi Sugiyama

Masashi Sugiyama (RIKEN-AIP/University of Tokyo, Japan) [homepage]
Coordinates: @ Zuse Institute Berlin (ZIB), Room 2006 (Seminar Room)

Title. Introduction of RIKEN Center for Advanced Intelligence Project

Abstract. (click to expand)

RIKEN is one of Japan’s largest fundamental-research institutions. The RIKEN Center for Advanced Intelligence Project (AIP) was created in 2016 to propose and investigate new machine learning methodologies and algorithms, and apply them to societal problems. AIP covers a wide range of topics from generic AI research (machine learning, optimization, applied math., etc.), goal-oriented AI research (material, disaster, cancer, etc.), and AI-in-society research (ethics, data circulation, laws, etc.). In the first half of my talk, I will briefly introduce our center’s activities and collaboration strategies.

Then, in the latter half, I will talk about the research activities in my team, i.e., machine learning from imperfect information. Machine learning has been successfully used to solve various real-world problems. However, we are still facing many technical challenges, for example, we want to train machine learning systems without big labeled data and we want to reliably deploy machine learning systems under noisy environments. I will overview our recent advances in tackling these problems.