Cognition  (C) Session 3

Abstract: Technology is a complex system that evolves through the collective efforts of individual inventors. Understanding inventors' behaviors may thus enable predicting invention or improving technology policy. We examined data from 2.8 million inventors' 4 million patents and found most patents are created by ``explorers": inventors who move across different technology domains during their careers. Explorers are far more likely to enter technology domains that were highly related to their own individual inventive experience; this information enabled accurate prediction of individual explorers' future movements. Inventors who entered very related domains patented more there, but explorers who successfully entered moderately related domains were more likely to create high-impact patents. These findings may be instructive for inventors exploring the space of technologies, and useful for organizations or governments in forecasting or directing technological change. Close

Abstract: We discuss the distribution of commuting distances and its relation to income. Using data from Denmark, the UK, and the US, we show that the commuting distance is (i) broadly distributed with a slow decaying tail that can be fitted by a power law with exponent γ ≈ 3 and (ii) an average growing slowly as a power law with an exponent less than one that depends on the country considered. The classical theory for job search is based on the idea that workers evaluate the wage of potential jobs as they arrive sequentially through time, and extending this model with space, we obtain predictions that are strongly contradicted by our empirical findings. We propose an alternative model that is based on the idea that workers evaluate potential jobs based on a quality aspect and that workers search for jobs sequentially across space. We also assume that the density of potential jobs depends on the skills of the worker and decreases with the wage. The predicted distribution of commuting distances decays as 1/r^3 and is independent of the distribution of the quality of jobs. We find our alternative model to be in agreement with our data. This type of approach opens new perspectives for the modeling of mobility. Close

Abstract: The well-known "Wisdom-of-Crowds" phenomenon, often mistakenly confused with collective intelligence, is not effective in every situation, especially under social influence. In simple estimation tasks, information sharing among group members may lead to strong biases in the collective estimate due to the reduction in diversity and independence of opinions. We are interested in finding conditions where social interactions could improve the accuracy of the collective estimate and its effectiveness. Specifying such conditions is an important step toward understanding how a human group can develop a form of collective intelligence emerging from social interactions between its members. We conduct a series of experiments aiming at understanding how a human group can use social information to converge toward the correct value in an estimation task. Subjects are sequentially asked to give a first guess, and then a second guess in the same estimation task after being provided with information about the average guess of the t previous subjects. We measure how this information affects the initial guesses of the subjects (with weight s) to various questions. We also measure the influence of "experts" (more knowledgeable subjects, introduced artificially with varying probability) and information lifetime (associated to t) on the convergence process. The distribution of social influence s is a Gaussian centered around s=2/3, with two additional very narrow peaks at 0 (highly confident subjects) and 1 ("followers"). We also find values of s below 0 or above 1, which correspond to subjects considered as "irrational" in micro-economic theories, and that may deeply affect the ability of a group to reach the right estimate. Unsurprisingly, the presence of experts improves both the final estimate and the speed of convergence. However, a decrease of the information lifetime t does not seem to influence the accuracy of the final estimate, but noticeably reduces the convergence time. Close

Abstract: We study the problem of event detection on distributed sensor networks. Prompt event detection is critical for many high-risk settings, for example evacuation following an earthquake. Distributed sensor networks are well suited for this task as they offer advantages such as high reliability and broad coverage. Distributed sensor networks can be organized in a centralized or decentralized way, the former offers higher accuracy, the latter lower communication volume. We study how this tradeoff varies for different network organizations, when a privacy cost is associated with communication. We assume that sensors pay a privacy cost for transmitting measurements. A real world example where this assumptions hold is earthquake detection with smartphones: If a device had to communicate regularly, the receiver could track its position throughout the day. We compare a centralized and a decentralized organization. In the centralized setting all sensors have to send their readings to the central event detection algorithm. In the decentralized setting the algorithm runs locally on a sensor, which alarms the central unit only if detects an event. This setting reduces the communication volume at the expense of accuracy. We propose a distributed protocol that reduces the privacy cost by reducing the communication to the central unit. Reducing communication is by itself desirable whenever remote communication is costly, for example low-power transmitters, and whenever computation is costly, for example if the central unit is a human supervisor with a limited attention span. The protocol allows sensors to ask their neighbors for an opinion on their measurements before reporting an event. The number of neighbors drives the accuracy/communication tradeoff. We test this protocol on different network topologies. We evaluate the system on detection accuracy and privacy cost. We expect to find a range of parameters for which a decentralized organization outperforms a centralized organization. Close

Abstract: People are constantly moving within cities and countries, facing the fact of the integration in habits and laws of new local cultures. Immigration phenomena have been studied and described so far by census data, which are indeed expensive to take, both in term of cost and time. Here we introduce a new methodology to explore the spatial integration of international immigrant communities in cities, exploring how Twitter users’ language might be a direct connection to their hometown and/or their nationality. We collect Twitter geo-localised data from 2012 to 2015 over a set of 58 out of the most populated cities in the world. We filter the users supposed to be residents in each city and their supposed-to-be place of residency. Finally, we assign to each user its most likely language. We conduct an extensive analysis on users’ spatial distribution within urban areas through a modified entropy metric, as a quantitative measure of the spatial integration of each language in the city. Results allowed us to characterized cities by their "Power of Integration”, as an attitude of hosting immigrant communities in urban areas, and by the corresponding process of integration of languages into different cultures, which is a quantitative measure of the differences between welcoming and hosting people in urban areas. Our findings provide a new way to detect the patterns of historically permanent immigration of people in urban areas, going beyond the estimation of past, current and foreshadowed global flows, towards a better comprehension of spatial integration phenomena on a city scale. Close