Panos Toulis

Friends with Faces: How Social Networks Can Enhance Face Recognition and Vice Versa

The “friendship” relation, a social relation among individuals, is one of the primary relations modeled in some of the world’s largest online social networking sites, such as “FaceBook.” On the other hand, the “co-occurrence” relation, as a relation among faces appearing in pictures, is one that is easily detectable using modern face detection techniques. These two relations, though appearing in different realms (social vs. visual sensory), have a strong correlation: faces that co-occur in photos often belong to individuals who are friends. Using real-world data gathered from “Facebook,” which were gathered as part of the “FaceBots” project, the world’s first physical face-recognizing and conversing robot that can utilize and publish information on “Facebook” was established. We present here methods as well as results for utilizing this correlation in both directions. Both algorithms for utilizing knowledge of the social context for faster and better face recognition are given, as well as algorithms for estimating the friendship network of a number of individuals given photos containing their faces. The results are quite encouraging. In the primary example, doubling of the recognition accuracy as well as a sixfold improvement in speed is demonstrated. Various improvements, interesting statistics, as well as an empirical investigation leading to predictions of scalability to much bigger data sets are discussed.

Scalable estimation strategies based on stochastic approximations: classical results and new insights

Estimation with large amounts of data can be facilitated by stochastic gradient methods, in which model parameters are updated sequentially using small batches of data at each step. Here, we review early work and modern results that illustrate the statistical properties of these methods, including convergence rates, stability, and asymptotic bias and variance. We then overview modern applications where these methods are useful, ranging from an online version of the EM algorithm to deep learning. In light of these results, we argue that stochastic gradient methods are poised to become benchmark principled estimation procedures for large datasets, especially those in the family of stable proximal methods, such as implicit stochastic gradient descent.

On the Synergies between Online Social Networking, Face Recognition and Interactive Robotics

This paper explores the intersection of three areas: interactive robots, face recognition, and online social networks, by presenting and discussing an implemented real-world system that combines all three, a “FaceBots” robot. Our robot is a mobile robot with face recognition, natural language dialogue, as well as mapping capabilities. The robot is also equipped with a social database containing information about the people it interacts with, and is also connected in real-time to the “Facebook” online social networking website, which contains information as well as partially tagged pictures. Our system demonstrates the benefits of this triangle of interconnection: it is not only the case that facebook information can lead to more interesting interactions, but also that: facebook photos enable better face recognition, interactive robots enable robot-mediated publishing of photos and information on facebook. Most importantly, as we shall see in detail, social information enables significantly better and faster face recognition, as an interesting bidirectional relationship exists between the “friends” relation in social networks and the “faces appear in the same picture” relation in face recognition. We will present algorithms for exploiting this relationship, as well as quantitative results. The two main novelties of our system are: this is the first interactive conversational mobile robot that utilizes and publishes social information in facebook, and is also the first system utilizing the social context of conjectured identities in a photo for better face recognition.

A long-term profit seeking strategy for continuous double auctions in a trading agent competition

This paper presents a new bidding strategy for continuous double auctions (CDA) designed for Mertacor, a successful trading agent, which won the first price in the “travel game” of Trading Agent Competition (TAC) for 2005. TAC provides a realistic benchmarking environment in which various travel commodities are offered in simultaneous online auctions. Among these, entertainment tickets are traded in CDA. The latter, represent the most dynamic part of the TAC game, in which agents are both sellers and buyers. In a CDA many uncertainty factors are introduced, because prices are constantly changing during the game and price fluctuations are hard to be predicted. In order to deal with these factors of uncertainty we have designed a strategy based on achieving a pre-defined long-term profit. This preserves the bidding attitude of our agent and shows flexibility in changes of the environment. We finally present and discuss the results of TAC-05, as well as an analysis of agents performance in the entertainment auctions.

Design and analysis of multi-hospital kidney exchange mechanisms using random graphs

Kidney exchanges enable transplants when a pair of a patient and an incompatible donor is matched with other similar pairs. In multi-hospital kidney exchanges pairs are pooled from multiple hospitals, and each hospital is able to decide which pairs to report and which to hide and match locally. Modeling the problem as a maximum matching on a random graph, we first establish that the expected benefit from pooling scales as the square-root of the number of pairs in each hospital. We design the xCM mechanism, which achieves efficiency and incentivizes hospitals of moderate-to-large size to fully report their pairs. Reciprocal pairs are crucial in the design, with the probabilistic uniform rule used to ensure incentive alignment. By grouping certain pair types into so-called virtual-reciprocal pairs, xCM extends to handle 3-cycles. We validate the performance of xCM in simulation, demonstrating its efficiency and incentive advantages over the Bonus mechanism (Ashlagi and Roth, 2014).

Incentive-Compatible Experimental Design

We consider the design of experiments to evaluate treatments that are administered by self-interested agents, each seeking to achieve the highest evaluation and win the experiment. For example, in an advertising experiment, a company wishes to evaluate two marketing agents in terms of their efficacy in viral marketing, and assign a contract to the winner agent. Contrary to traditional experimental design, this problem has two new implications. First, the experiment induces a game among agents, where each agent can select from multiple versions of the treatment it administers. Second, the action of one agent -- selection of treatment version -- may affect the actions of another agent, with the resulting strategic interference complicating the evaluation of agents. An incentive-compatible experiment design is one with an equilibrium where each agent selects its natural action, which is the action that maximizes the performance of the agent without competition (e.g., expected number of conversions if agent is assigned the advertising contract). Under a general formulation of block experiment designs, we identify sufficient conditions that guarantee incentive-compatible experiments.These conditions rely on the existence of statistics that can estimate how agents would perform without competition,and their use in constructing score functions to evaluate the agents. In the setting with no strategic interference, we also study the power of the design, i.e., the probability that the best agent wins, and show how to improve the power of incentive-compatible designs.From the technical side, our theory uses a range of statistical methods such as hypothesis testing, variance-stabilizing transformations and the Delta method, all of which rely on asymptotics.

Stochastic Gradient Methods for Principled Estimation with Large Datasets

14.

Asymptotic and finite-sample properties of estimators based on stochastic gradients

Stochastic gradient descent procedures have gained popularity for parameter estimation from large data sets. However, their statistical properties are not well understood, in theory. And in practice, avoiding numerical instability requires careful tuning of key parameters. Here, we introduce implicit stochastic gradient descent procedures, which involve parameter updates that are implicitly defined. Intuitively, implicit updates shrink standard stochastic gradient descent updates. The amount of shrinkage depends on the observed Fisher information matrix, which does not need to be explicitly computed; thus, implicit procedures increase stability without increasing the computational burden. Our theoretical analysis provides the first full characterization of the asymptotic behavior of both standard and implicit stochastic gradient descent-based estimators, including finite-sample error bounds. Importantly, analytical expressions for the variances of these stochastic gradient-based estimators reveal their exact loss of efficiency. We also develop new algorithms to compute implicit stochastic gradient descent-based estimators for generalized linear models, Cox proportional hazards, M-estimators, in practice, and perform extensive experiments. Our results suggest that implicit stochastic gradient descent procedures are poised to become a workhorse for approximate inference from large data sets