Upcoming events

Fairness considerations in settings where decisions are made by supervised learning algorithms (e.g. criminal risk assessment) has received considerable attention, recently. As the fairness literature continues to expand mostly around this canonical learning task, it is important to recognize that many real-world applications of ML fall outside the category of supervised, one-shot learning. In this presentation, I will talk about two scenarios in which algorithmic decisions are made in a sequential manner and over time. I will argue that in such settings, being fair---at a minimum---requires decisions to be "consistent" across individuals who arrive at different time steps, that is, similar individuals must be treated similarly. I will then talk about how such consistency constraints affect learning. 

In the first part of the talk, I will introduce a generic sequential decision making framework, in which at each time step the learning algorithm receives data corresponding to a new individual (e.g. a new job application) and must make an irrevocable decision about him/her (e.g. whether to hire the applicant) based on observations it has made so far. I propose a general framework for post-processing predictions made by a black-box learning model, so that the resulting sequence of outcomes is guaranteed to be consistent. I show, both theoretically and via simulations, that imposing consistency constraints will not significantly slow down learning.

In the second part of the talk, I will focus on fairness considerations in a particular type of market---namely, combinatorial prediction markets---where traders can submit limit orders on various security bundles, and the market maker is tasked with executing these orders in a fair manner. The main challenge in running such market is that executing one order can potentially change the price of every other order in the book. I define the notion of a "fair trading path", which at a high level guarantees similar orders are executed similarly: no order executes at a price above its limit, and every order executes when its market price falls below its limit price. I present a market algorithm that respects these fairness conditions, and evaluate it using real combinatorial predictions made during the 2008 U.S. Presidential election. 

I will conclude by comparing my work with previous papers on fairness for online learning, and a list of directions for future work.

A variety of experts -- computer scientists, policy makers, judges -- constantly make decisions about best practices for computational systems. They decide which features are fair to use in a machine learning classifier predicting whether someone will commit a crime, and which security behaviors to recommend and require from end-users. Yet, the best decision is not always clear. Studies have shown that experts often disagree with each other and, perhaps more importantly, with the people for whom they are making these decisions: the users.

This raises a question: Is it possible to learn best practices directly from the users? The field of moral philosophy suggests yes, through the process of descriptive decision-making, in which we observe people's preferences and then infer best practice rather than using experts' normative (prescriptive) determinations to define best practice. In this talk, I will explore the benefits and challenges of applying such a descriptive approach to making computationally relevant decisions regarding: (i) selecting security prompts for an online system; (ii) determining which features to include in a classifier for jail sentencing; (iii) defining standards for ethical virtual reality content.

SCION is a new Internet architecture that addresses many of the security vulnerabilities of today's Internet. Its clean-slate design provides, among other properties, route control, failure isolation, and multi-path communication. The verifiedSCION project is an effort to formally verify the correctness and security of SCION. It aims to provide strong guarantees for the entire architecture, from the protocol design to its concrete implementation. The project uses stepwise refinement to prove that the protocol withstands increasingly strong attackers. The refinement proofs assume that all network components such as routers satisfy their specifications. This property is then verified separately using deductive program verification in separation logic. This talk will give an overview of the verifiedSCION project and explain, in particular, how we verify code-level properties such as memory safety, I/O behavior, and information flow security.

In real-time embedded systems, for each task the compliance to timing constraints has to be guaranteed in addition to the functional correctness. The first part of the talk considers the theoretical comparison of scheduling algorithms and schedulability tests by evaluating speedup factors for non-preemptive scheduling, which leads to a discussion about general problems of resource augmentation bounds. In addition, it is explained how utilization bounds can be parametrized, resulting in better bounds for specific scenarios, i.e., when analyzing non-preemptive Rate-Monotonic scheduling as well as task sets inspired by automotive applications.

In the second part, a setting similar to mixed-criticality systems is considered and the criticism on previous work in this area is detailed. Hence, a new system model that allows a better applicability to realistic scenarios, namely Systems with Dynamic Real-Time Guarantees, is explained. This model is extended to a multiprocessor scenario, considering CPU overheating as a possible cause for mixed-criticality behaviour. Finally, a way to determine the deadline-miss probability for such systems is described that drastically reduces the runtime of such calculations.

The third part discusses tasks with self-suspension behaviour, explains a fixed-relative-deadline strategy for segmented self-suspension tasks with one suspension interval, and details how this approach can be exploited in a resource-oriented partitioned scheduling. Furthermore, it is explained how the gap between the dynamic and the segmented self-suspension model can be bridged by hybrid models.

I will present a recent non-elementary lower bound for the complexity of reachability problem for Vector Addition Systems. I plan to show the main insights of the proof. In particular I will present a surprising equation on fractions, which is the core of the new source of hardness found in VASes.

Many classification tasks have no objective ground truth. Examples include: which content or explanation is "better" according to some community? is this comment toxic? what is the political leaning of this news article? The traditional modeling approach assumes each item has an objective true state that is perceived by humans with some random error. It fails to account for the fact that people have greater agreement on some items than others. I will describe an alternative model where the true state is a distribution over labels that raters from a specified population would assign to an item. This leads to information gain (mutual information) as a theoretically justified and computationally tractable measure of a classifier's quality, and an intuitive interpretation of information gain in terms of the sample size for a survey that would yield the same expected error rate.

tbd

tbd