NLP in the era of generative AI, cognitive sciences, and societal transformation
Join us at Mila in October for a three-day workshop to explore the transformative potential of language technologies and their implications for society.
This program is designed to provide decision-makers, policymakers and professional working in policy with a foundational understanding of AI technology.
We use cookies to analyze the browsing and usage of our website and to personalize your experience. You can disable these technologies at any time, but this may limit certain functionalities of the site. Read our Privacy Policy for more information.
Setting cookies
You can enable and disable the types of cookies you wish to accept. However certain choices you make could affect the services offered on our sites (e.g. suggestions, personalised ads, etc.).
Essential cookies
These cookies are necessary for the operation of the site and cannot be deactivated. (Still active)
Analytics cookies
Do you accept the use of cookies to measure the audience of our sites?
Multimedia Player
Do you accept the use of cookies to display and allow you to watch the video content hosted by our partners (YouTube, etc.)?
Publications
Automated, interactive, and traceable domain modelling empowered by artificial intelligence
Estimating value functions is a core component of reinforcement learning algorithms. Temporal difference (TD) learning algorithms use bootst… (see more)rapping, i.e. they update the value function toward a learning target using value estimates at subsequent time-steps. Alternatively, the value function can be updated toward a learning target constructed by separately predicting successor features (SF)—a policy-dependent model—and linearly combining them with instantaneous rewards.
We focus on bootstrapping targets used when estimating value functions, and propose a new backup target, the ?-return mixture, which implicitly combines value-predictive knowledge (used by TD methods) with (successor) feature-predictive knowledge—with a parameter ? capturing how much to rely on each. We illustrate that incorporating predictive knowledge through an ??-discounted SF model makes more efficient use of sampled experience, compared to either extreme, i.e. bootstrapping entirely on the value function estimate, or bootstrapping on the product of separately estimated successor features and instantaneous reward models. We empirically show this approach leads to faster policy evaluation and better control performance, for tabular and nonlinear function approximations, indicating scalability and generality.
BACKGROUND
Robust and flexible infectious disease surveillance is crucial for public health. Event-based surveillance (EBS) was developed t… (see more)o allow timely detection of infectious disease outbreaks using mostly web-based data. Despite its widespread use, EBS has not been evaluated systematically on a global scale in terms of outbreak detection performance.
OBJECTIVE
To assess the variation in timing and frequency of EBS reports compared to true outbreaks and to identify the determinants of variability, using the example of seasonal influenza epidemics in 24 countries.
METHODS
We obtained influenza-related reports from two EBS systems, HealthMap and the WHO Epidemic Intelligence from Open Sources (EIOS), and weekly virologic influenza counts from FluNet as a gold standard. Epidemic influenza periods were detected based on report frequency using Bayesian change point analysis. Timely sensitivity, i.e., outbreak detection within the first two weeks before or after an outbreak onset, was calculated along with sensitivity, specificity, positive predictive value, and timeliness of detection. Linear regressions were performed to assess the influence of country-specific factors on EBS performance.
RESULTS
Overall, monitoring the frequency of EBS reports detected 73.5% of outbreaks, but only 9.2% within two weeks of onset; in the best case, monitoring the frequency of health-related reports identified 29% of outbreaks within two weeks of onset. We observed a large degree of variability in all evaluation metrics across countries. The number of EBS reports available within a country, the human development index, and the country’s geographical location partially explained this variability.
CONCLUSIONS
Monitoring the frequency of EBS reports allowed just under 10% of seasonal influenza outbreaks to be detected in a timely manner in a worldwide analysis, with a large variability in detection capabilities. This article documents the global variation of EBS performance and demonstrates that monitoring report frequency alone in EBS may be insufficient for timely detection of outbreaks. Moreover, factors such as human development index and geographical location of a country may influence the performance of EBS and should be considered in future evaluations.
BACKGROUND
Robust and flexible infectious disease surveillance is crucial for public health. Event-based surveillance (EBS) was developed t… (see more)o allow timely detection of infectious disease outbreaks using mostly web-based data. Despite its widespread use, EBS has not been evaluated systematically on a global scale in terms of outbreak detection performance.
OBJECTIVE
To assess the variation in timing and frequency of EBS reports compared to true outbreaks and to identify the determinants of variability, using the example of seasonal influenza epidemics in 24 countries.
METHODS
We obtained influenza-related reports from two EBS systems, HealthMap and the WHO Epidemic Intelligence from Open Sources (EIOS), and weekly virologic influenza counts from FluNet as a gold standard. Epidemic influenza periods were detected based on report frequency using Bayesian change point analysis. Timely sensitivity, i.e., outbreak detection within the first two weeks before or after an outbreak onset, was calculated along with sensitivity, specificity, positive predictive value, and timeliness of detection. Linear regressions were performed to assess the influence of country-specific factors on EBS performance.
RESULTS
Overall, monitoring the frequency of EBS reports detected 73.5% of outbreaks, but only 9.2% within two weeks of onset; in the best case, monitoring the frequency of health-related reports identified 29% of outbreaks within two weeks of onset. We observed a large degree of variability in all evaluation metrics across countries. The number of EBS reports available within a country, the human development index, and the country’s geographical location partially explained this variability.
CONCLUSIONS
Monitoring the frequency of EBS reports allowed just under 10% of seasonal influenza outbreaks to be detected in a timely manner in a worldwide analysis, with a large variability in detection capabilities. This article documents the global variation of EBS performance and demonstrates that monitoring report frequency alone in EBS may be insufficient for timely detection of outbreaks. Moreover, factors such as human development index and geographical location of a country may influence the performance of EBS and should be considered in future evaluations.
Background Robust and flexible infectious disease surveillance is crucial for public health. Event-based surveillance (EBS) was developed to… (see more) allow timely detection of infectious disease outbreaks by using mostly web-based data. Despite its widespread use, EBS has not been evaluated systematically on a global scale in terms of outbreak detection performance. Objective The aim of this study was to assess the variation in the timing and frequency of EBS reports compared to true outbreaks and to identify the determinants of variability by using the example of seasonal influenza epidemic in 24 countries. Methods We obtained influenza-related reports between January 2013 and December 2019 from 2 EBS systems, that is, HealthMap and the World Health Organization Epidemic Intelligence from Open Sources (EIOS), and weekly virological influenza counts for the same period from FluNet as the gold standard. Influenza epidemic periods were detected based on report frequency by using Bayesian change point analysis. Timely sensitivity, that is, outbreak detection within the first 2 weeks before or after an outbreak onset was calculated along with sensitivity, specificity, positive predictive value, and timeliness of detection. Linear regressions were performed to assess the influence of country-specific factors on EBS performance. Results Overall, while monitoring the frequency of EBS reports over 7 years in 24 countries, we detected 175 out of 238 outbreaks (73.5%) but only 22 out of 238 (9.2%) within 2 weeks before or after an outbreak onset; in the best case, while monitoring the frequency of health-related reports, we identified 2 out of 6 outbreaks (33%) within 2 weeks of onset. The positive predictive value varied between 9% and 100% for HealthMap and from 0 to 100% for EIOS, and timeliness of detection ranged from 13% to 94% for HealthMap and from 0% to 92% for EIOS, whereas system specificity was generally high (59%-100%). The number of EBS reports available within a country, the human development index, and the country’s geographical location partially explained the high variability in system performance across countries. Conclusions We documented the global variation of EBS performance and demonstrated that monitoring the report frequency alone in EBS may be insufficient for the timely detection of outbreaks. In particular, in low- and middle-income countries, low data quality and report frequency impair the sensitivity and timeliness of disease surveillance through EBS. Therefore, advances in the development and evaluation and EBS are needed, particularly in low-resource settings.
The epidemiological impact of the Canadian COVID Alert app
Shuo Sun
Mairead Shaw
Erica E.M. Moodie
Derek Ruths
We analyzed the effectiveness of the Canadian COVID Alert app on reducing COVID-19 infections and deaths due to the COVID-19 virus. Two sepa… (see more)rate but complementary approaches were taken. First, we undertook a comparative study to assess how the adoption and usage of the COVID Alert app compared to those of similar apps deployed in other regions. Next, we used data from the COVID Alert server and a range of plausible parameter values to estimate the numbers of infections and deaths averted in Canada using a model that combines information on number of notifications, secondary attack rate, expected fraction of transmissions that could be prevented, quarantine effectiveness, and expected size of the full transmission chain in the absence of exposure notification. The comparative analysis revealed that the COVID Alert app had among the lowest adoption levels among apps that reported usage. Our model indicates that use of the COVID Alert app averted between 6284 and 10,894 infections across the six Canadian provinces where app usage was highest during the March–July 2021 period. This range is equivalent to 1.6–2.9% of the total recorded infections across Canada in that time. Using province-specific case fatality rates, 57–101 deaths were averted during the same period. The number of cases and deaths averted was greatest in Ontario, whereas the proportion of cases and deaths averted was greatest in Newfoundland and Labrador. App impact measures were reported so rarely and so inconsistently by other regions that the relative assessment of impact is inconclusive. While the nationwide rates are low, provinces with widespread adoption of the app showed high ratios of averted cases and deaths (upper bound was greater than 60% of averted cases). This finding suggests that the COVID Alert app, when adopted at sufficient levels, can be an effective public health tool for combatting a pandemic such as COVID-19.
The epidemiological impact of the Canadian COVID Alert app
Shuo Sun
Mairead Shaw
Erica E. M. Moodie
Derek Ruths
We analyzed the effectiveness of the Canadian COVID Alert app on reducing COVID-19 infections and deaths due to the COVID-19 virus. Two sepa… (see more)rate but complementary approaches were taken. First, we undertook a comparative study to assess how the adoption and usage of the COVID Alert app compared to those of similar apps deployed in other regions. Next, we used data from the COVID Alert server and a range of plausible parameter values to estimate the numbers of infections and deaths averted in Canada using a model that combines information on number of notifications, secondary attack rate, expected fraction of transmissions that could be prevented, quarantine effectiveness, and expected size of the full transmission chain in the absence of exposure notification. The comparative analysis revealed that the COVID Alert app had among the lowest adoption levels among apps that reported usage. Our model indicates that use of the COVID Alert app averted between 6284 and 10,894 infections across the six Canadian provinces where app usage was highest during the March–July 2021 period. This range is equivalent to 1.6–2.9% of the total recorded infections across Canada in that time. Using province-specific case fatality rates, 57–101 deaths were averted during the same period. The number of cases and deaths averted was greatest in Ontario, whereas the proportion of cases and deaths averted was greatest in Newfoundland and Labrador. App impact measures were reported so rarely and so inconsistently by other regions that the relative assessment of impact is inconclusive. While the nationwide rates are low, provinces with widespread adoption of the app showed high ratios of averted cases and deaths (upper bound was greater than 60% of averted cases). This finding suggests that the COVID Alert app, when adopted at sufficient levels, can be an effective public health tool for combatting a pandemic such as COVID-19.
Self-Supervised Learning (SSL) with large-scale unlabelled datasets enables learning useful representations for multiple downstream tasks. H… (see more)owever, assessing the quality of such representations efficiently poses nontrivial challenges. Existing approaches train linear probes (with frozen features) to evaluate performance on a given task. This is expensive both computationally, since it requires retraining a new prediction head for each downstream task, and statistically, requires task-specific labels for multiple tasks. This poses a natural question, how do we efficiently determine the "goodness" of representations learned with SSL across a wide range of potential downstream tasks? In particular, a task-agnostic statistical measure of representation quality, that predicts generalization without explicit downstream task evaluation, would be highly desirable.
In this work, we analyze characteristics of learned representations
We propose a theoretical framework for approximate planning and learning in partially observed systems. Our framework is based on the fundam… (see more)ental notion of information state. We provide two equivalent definitions of information state---i) a function of history which is sufficient to compute the expected reward and predict its next value; ii) equivalently, a function of the history which can be recursively updated and is sufficient to compute the expected reward and predict the next observation. An information state always leads to a dynamic programming decomposition. Our key result is to show that if a function of the history (called approximate information state (AIS)) approximately satisfies the properties of the information state, then there is a corresponding approximate dynamic program. We show that the policy computed using this is approximately optimal with bounded loss of optimality. We show that several approximations in state, observation and action spaces in literature can be viewed as instances of AIS. In some of these cases, we obtain tighter bounds. A salient feature of AIS is that it can be learnt from data. We present AIS based multi-time scale policy gradient algorithms. and detailed numerical experiments with low, moderate and high dimensional environments.