No abstract

Abstract-We draw a formal connection between using synthetic training data to optimize neural network parameters and approximate, Bayesian, model-based reasoning. In particular, training a neural network using synthetic data can be viewed as learning a proposal distribution generator for approximate inference in the synthetic-data generative model. We demonstrate this connection in a recognition task where we develop a novel Captcha-breaking architecture and train it using synthetic data, demonstrating both state-of-the-art performance and a way of computing task-specific posterior uncertainty. Using a neural network trained this way, we also demonstrate successful breaking of real-world Captchas currently used by Facebook and Wikipedia. Reasoning from these empirical results and drawing connections with Bayesian modeling, we discuss the robustness of synthetic data results and suggest important considerations for ensuring good neural network generalization when training with synthetic data.

Significance We resolve conflicting results regarding mask wearing against COVID-19. Most previous work focused on mask mandates; we study the effect of mask wearing directly. We find that population mask wearing notably reduced SARS-CoV-2 transmission (mean mask-wearing levels corresponding to a 19% decrease in R). We use the largest wearing survey (n = 20 million) and obtain our estimates from regions across six continents. We account for nonpharmaceutical interventions and time spent in public, and quantify our uncertainty. Factors additional to mask mandates influenced the worldwide early uptake of mask wearing. Our analysis goes further than past work in the quality of wearing data–100 times the size with random sampling–geographical scope, a semimechanistic infection model, and the validation of our results.

Word embeddings obtained from neural network models such as Word2Vec Skipgram have become popular representations of word meaning and have been evaluated on a variety of word similarity and relatedness norming data. Skipgram generates a set of word and context embeddings, the latter typically discarded after training. We demonstrate the usefulness of context embeddings in predicting asymmetric association between words from a recently published dataset of production norms (Jouravlev and McRae, 2016). Our findings suggest that humans respond with words closer to the cue within the context embedding space (rather than the word embedding space), when asked to generate thematically related words.

Mask-wearing has been a controversial measure to control the COVID-19 pandemic. While masks are known to substantially reduce disease transmission in healthcare settings (Howard et al 2021), studies in community settings report inconsistent results (Brainard et al 2020). Investigating the inconsistency within epidemiological studies, we find that a commonly used proxy, government mask mandates, does not correlate with large increases in mask-wearing in our window of analysis. We thus analyse the effect of mask-wearing on transmission instead, drawing on several datasets covering 92 regions on 6 continents, including the largest survey of individual-level wearing behaviour (n=20 million) (Kreuter et al 2020). Using a hierarchical Bayesian model, we estimate the effect of both mask-wearing and mask-mandates on transmission by linking wearing levels (or mandates) to reported cases in each region, adjusting for mobility and non-pharmaceutical interventions. We assess the robustness of our results in 123 experiments across 22 sensitivity analyses. Across these analyses, we find that an entire population wearing masks in public leads to a median reduction in the reproduction number R of 25.8%, with 95% of the medians between 22.2% and 30.9%. In our window of analysis, the median reduction in $R$ associated with the wearing level observed in each region was 20.4% [2.0%, 23.3%]. We do not find evidence that mandating mask-wearing reduces transmission. Our results suggest that mask-wearing is strongly affected by factors other than mandates. We establish the effectiveness of mass mask-wearing, and highlight that wearing data, not mandate data, are necessary to infer this effect.