OVERVIEW OF THE INDUSTRY PROGRAM

Modeling air pollution: Informing policies to address a global environmental challenge

Air pollution is a leading cause of global mortality: according to the World Health Organization, 90% of people worldwide breathe polluted air, and outdoor air pollution causes over 4 million deaths annually. Different models can simulate various aspects of the air pollution problem by quantifying pollutant emissions from different sectors and their socio-economic drivers, tracing the chemistry and transport of atmospheric processes, attributing pollutant concentrations to specific sources, and quantifying the health and economic burdens of pollutant exposure. To inform efforts to mitigate air pollution, however, we need to trace the entire pathway by which policies to address emissions translate into societal benefits. Doing this requires connecting models from different academic fields, and which exist in different modeling languages, with different temporal and spatial scales, and with different core scientific assumptions. In this talk, I summarize work from my research group evaluating the impacts of air pollution policies by connecting and integrating models across this conceptual chain. Examples provided include assessing the air pollution and related health impacts of proposed policies to address climate change in the U.S. and China, and quantifying the domestic and international benefits of mercury reduction policies in China, India, and the U.S. Technical challenges of linking models include accounting for issues of temporal and spatial scale, complexity, and boundaries. Effectively informing decision-making, however, also requires that decision-makers see models as credible, salient, and legitimate. Thus, I also describe ways in which we have engaged with stakeholders and decision-makers, and examine how these efforts have influenced the impact of this research on policy.

Models for motion prediction; robotic brains versus biological brains

This talk will focus on robot and animal motions. For any system in the real world (robots, animals, and humans alike), for any task that they need to execute, the world needs to be predictable. Accurate predictions facilitate anticipation, planning, and optimization. And predictions are made with a model, i.e. a description of (relevant parts of) the world written in a well-defined language. So the key question for this talk is which models and which modelling language are used by robots and by the biological brain, and is there any similarity between the two? We will first review which models and which language are used for the prediction of dynamic motions for robots. We will survey the basics of (systems of) differential equations illustrated with our work on walking robots, self-driving cars, robotic arms, hands, and integrated systems for logistic applications (the "Amazon Picking Challenge"). Next, in this presentation we survey the reigning neuroscientific theories for how the brain does modelling and prediction. Prediction will appear to play a deep fundamental role in all of our brain functions, and we will analyse how this is related to the robot approach; the brain has no processing unit for matrix multiplications or differential equations, so how does the brain do dynamic prediction? Finally, we will share a preliminary breakthrough result termed "Dualistic Prediction Error Minimization", a key insight from the brain's ways of prediction error minimization, which will allow for a tremendous improvement in modelling accuracy for our future complex robot systems that require predictability within a complex environment.

Genetics of Computer Programs

Language use vectors (LUVs) encode the vocabulary and frequency of programming language feature use by computer programs and applications. In some sense, they encode the potential and expressed vocabulary of a program in a way analogous to the way that the genome (DNA) and phenome (RNA) describe the potential and expressed properties of a living thing. In this talk I will explore this analogy and its potential for predicting hidden properties of programs, models and applications such as changes, faults and security flaws.