No abstract

We demonstrate that a production-quality keyword-spotting model can be trained on-device using federated learning and achieve comparable false accept and false reject rates to a centrally-trained model. To overcome the algorithmic constraints associated with fitting on-device data (which are inherently non-independent and identically distributed), we conduct thorough empirical studies of optimization algorithms and hyperparameter configurations using large-scale federated simulations. To overcome resource constraints, we replace memoryintensive MTR data augmentation with SpecAugment, which reduces the false reject rate by 56%. Finally, to label examples (given the zero visibility into on-device data), we explore teacher-student training.

Successive waves of infection by SARS-CoV-2 have left little doubt that this virus will transition to an endemic disease 1,2. Projections of the endemic seasonality of SARS-CoV-2 transmission are crucial to informed public health policy 3. Such projections are not only essential to well-timed interventions and the preparation of healthcare systems for synchronous surges with other respiratory viruses 4, but also to the elimination of seasonality as a confounder in the identification of surges that are occurring due to viral evolution, changes in host immunity, or other non-seasonal factors. However, the less than two-year duration of SARS-CoV-2 circulation, pandemic dynamics, and heterogeneous implementation of interventions have grievously complicated evaluations of its seasonality 5. Here we estimate the impending endemic seasonality of SARS-CoV-2 in global population centers via a novel phylogenetic ancestral and descendent states approach 6 that leverages long-term data on the incidence of circulating coronaviruses. Our results validate a major concern that endemic COVID-19 will typically surge coincident with other high-morbidity and -mortality respiratory virus infections such as influenza and RSV 7. In temperate locales in the Northern Hemisphere, we identify spatiotemporal surges of incidences that range from October through January in New York to January through March in Yamagata, Japan. This knowledge of likely spatiotemporal surges of COVID-19 is fundamental to optimal timing of public health interventions that anticipate the impending endemicity of this disease and mitigate SARS-CoV-2 transmission.

Background & Aims: Human genetic variation is thought to guide the outcome of hepatitis C virus (HCV) infection but model systems within which to dissect these host genetic mechanisms are limited. Norway rat hepacivirus (NrHV), closely related to HCV, causes chronic liver infection in rats but causes acute self-limiting hepatitis in typical strains of laboratory mice, which resolves in two weeks. The Collaborative Cross (CC) is a robust mouse genetics resource comprised of a panel of recombinant inbred strains, which model the complexity of the human genome and provide a system within which to understand diseases driven by complex allelic variation. Approach & Results: We infected a panel of CC strains with NrHV and identified several that failed to clear virus after 4 weeks. Strains displayed an array of virologic phenotypes ranging from delayed clearance (CC046) to chronicity (CC071, CC080) with viremia for at least 10 months. Body weight loss, hepatocyte infection frequency, viral evolution, T-cell recruitment to the liver, liver inflammation and the capacity to develop liver fibrosis varied among infected CC strains. Conclusions: These models recapitulate many aspects of HCV infection in humans and demonstrate that host genetic variation affects a multitude of virus and host phenotypes. These models can be used to better understand the molecular mechanisms that drive hepacivirus clearance and chronicity, the virus and host interactions that promote chronic disease manifestations like liver fibrosis, therapeutic and vaccine performance, and how these factors are affected by host genetic variation.

Background & Aims: Human genetic variation is thought to guide the outcome of hepatitis C virus (HCV) infection but model systems within which to dissect these host genetic mechanisms are limited. Norway rat hepacivirus (NrHV), closely related to HCV, causes chronic liver infection in rats but causes acute self-limiting hepatitis in typical strains of laboratory mice, which resolves in two weeks. The Collaborative Cross (CC) is a robust mouse genetics resource comprised of a panel of recombinant inbred strains, which model the complexity of the human genome and provide a system within which to understand diseases driven by complex allelic variation. Approach & Results: We infected a panel of CC strains with NrHV and identified several that failed to clear virus after 4 weeks. Strains displayed an array of virologic phenotypes ranging from delayed clearance (CC046) to chronicity (CC071, CC080) with viremia for at least 10 months. Body weight loss, hepatocyte infection frequency, viral evolution, T-cell recruitment to the liver, liver inflammation and the capacity to develop liver fibrosis varied among infected CC strains. Conclusions: These models recapitulate many aspects of HCV infection in humans and demonstrate that host genetic variation affects a multitude of virus and host phenotypes. These models can be used to better understand the molecular mechanisms that drive hepacivirus clearance and chronicity, the virus and host interactions that promote chronic disease manifestations like liver fibrosis, therapeutic and vaccine performance, and how these factors are affected by host genetic variation.