Weizmann Institute Interactive Map

Program

14:00 Guillermo Sapiro, Duke University and Apple

Applying computer vision and machine learning to a truly hard problem: Developmental health

14:30 Avrim Blum, Toyota Technological Institute at Chicago (TTIC)

On learning in the presence of biased data and strategic behavior

15:00 Bill Freeman, MIT

Finding Pareidolic Andrew

15:30 Coffee Break

16:00 Leonard Schulman, Caltech

Mixture Models and Causal Inference

16:30 Allan Borodin, University of Toronto

Two Voting Issues (Gerrymandering and Primaries): Can Theoretical Computer Science Provide Any Insights?

Organizer: Moni Naor

Contact for information: Sharon.Rom@weizmann.ac.il

Abstracts

Guillermo Sapiro: Applying computer vision and machine learning to a truly hard problem: Developmental health

In this work we will describe how we deployed computer vision and machine learning tools to address developmental health challenges in real world settings, from pediatric clinics to homes. We will illustrate the effort, its challenges, and achievements, which range from replicating lab-based work with consumer grade tools to discovering new biomarkers for autism.

Avrim Blum: On learning in the presence of biased data and strategic behavior

In this talk I will discuss two lines of work involving learning in the presence of biased data and strategic behavior. In the first, we ask whether fairness constraints on learning algorithms can actually improve the accuracy of the classifier produced, when training data is unrepresentative or corrupted due to bias. Typically, fairness constraints are analyzed as a tradeoff with classical objectives such as accuracy. Our results here show there are natural scenarios where they can be a win-win, helping to improve overall accuracy. In the second line of work we consider strategic classification: settings where the entities being measured and classified wish to be classified as positive (e.g., college admissions) and will try to modify their observable features if possible to make that happen. We consider this in the online setting where a particular challenge is that updates made by the learning algorithm will change how the inputs behave as well.

Bill Freeman: Finding Pareidolic Andrew

This is a talk I gave at a workshop celebrating Andrew Zisserman. Pareidolia is the human ability to see shapes or make pictures out of randomness. After motivating the task--finding pareidolic Andrew--we introduce a simple model, which predicts, under Gaussian assumptions, conditions favorable for pareidolia, predicting "peak pareidolia". A psychophysics survey gives support to the findings of the simple model. Armed with those results, we seek pareidolic images of Andrew in (a) contours, and (b) 2-d images.

Leonard Schulman: Mixture Models and Causal Inference

Bayesian networks model direct causal effects among variables. A fundamental problem in Causal Inference is to identify total causal effects despite the existence of unobservable confounding variables. Sometimes this is possible due to specific structure of the network; but generally the structure is too weak to allow this. In that case another condition which may render the task feasible is a cardinality bound on the range of confounders. We have been studying this direction in a series of papers, which build up toward this goal through the key problem of identifying mixtures of product distributions. I'll survey a few points from our work.

Based on joint works with C. Swamy (Waterloo), Y. Rabani (Hebrew U), J. Li (Tsinghua), S. Gordon (Caltech) and B. Mazaheri (Caltech)

Allan Borodin: Two Voting Issues (Gerrymandering and Primaries): Can Theoretical Computer Science Provide Any Insights?

Gerrymandering is the process where a given partisan body (i.e., a political party) can form voting districts so as to obtain a significant advantage in the number of ``seats'' won. We model this problem in graph theoretic terms and provide strong evidence that even without any gerrymandering there is some inherent bias toward the ``rural party''. Gerrymandering can be done computationally (without human expertise) by any party that controls the districting process to produce outcomes that are highly non representative of general voter preferences. Our evidence is thus far of an experimental nature based on both synthetic data and actual voting data. Our model provides opportunities for understanding possible future voting outcomes.

Understanding the ``rural-urban'' divide is of interest to any voting system based on district voting.

Voting with candidates chosen by primaries are best known in US elections. But the phenomenon of multi-stage elections appears in other contexts.

We study the impact of primaries with respect to final outcomes (in the general election) using the framework of metric distortion. Here we model voter preferences for candidates by their distance in a metric space which is used to define the political views of voters and candidates.

With respect to primaries we can obtain a number of theoretical results in terms of distortion. We further create synthetic data to better understand the distortion effect of primaries on the quality of the final election winner.

Joint work with Omer lev, Nisarg Shah and Tyrone Strangway