Skip to content

Home

Jump to bottom
isaacmg edited this page Jun 28, 2020 · 42 revisions

Task Time Series (TS)

Cross Roads

Task TS sits at the crossroads of a lot of interesting areas within epidemiology, machine learning, time-series forecasting, data analysis, and software engineering. We aim to provide value to the broader CoronaWhy organization and global AI4Good community in the following ways:

(1) Coronavirus and response

  • Develop models that can effectively forecast new Coronavirus cases, deaths, and hospitalizations 14 days into the future. These can aid in public-policy decisions about when to look-down or open up specific counties. They can also inform hospitals of when to stock up on PPE.
  • Use trained models to provide insights into virus transmission mechanics and utility of public policy interventions.
  • Develop a central publicly available temporal data lake for Coronavirus that other epidemiologists and researchers can leverage.
  • Develop models to help forecast the economic impact in terms of jobs and GDP of proposed interventions.
  • If possible develop models to forecast patient length of stay in hospitals and risk of decompensation to aid in discharge planning and early warning systems for physicians.

(2) Epidemiology

  • Demonstrate the utility of temporal machine learning as a broader tool for studying pandemics, spread of disease, and public health crises.
  • Show how temporal machine learning models can help gauge the impact of public policy interventions.

(3) Machine Learning Research

  • Develop few shot learning techniques for time series forecasting that can effectively train models to forecast on limited data.
  • Devise multi-modal learning techniques that can effectively incorporate geo-spatial data and meta-data with temporal data.
  • Cracking one the "black-box" and providing new ways to explain deep models.
  • Figure out innovative ways to integrate traditional statistical time series methods with latest in DL research from NLP and CV.

(4) ML Ops

  • Showcase experiment tracking and extendibility best practices with configuration files.
  • Show template for proper unit tests and continuous deployment of packages and ML models to production.
  • Demonstrate how versioned datasets enable completely reproducible results.
  • Develop continuous retraining pipelines that automatically train and re-deploy top models as new data comes in.

(5) Data Science Teams

Goals/Concrete Deliverables

  1. Publications in top Machine Learning Conferences on novel time series forecasting models and few-shot learning methods

  2. Publications in top epidemiology journals on machine learning as methodological tool in studying COVID-19 and recommended interventions based on our model.

  3. Daily dashboard of expected Coronavirus cases, deaths, hospitalizations in all U.S. and Western European Counties with contributing factors listed.

  4. Task agnostic forecasting framework that can be leveraged to forecast time series problem.

  5. Geo-spatial/temporal data lake of COVID-19, SARS, MERs, Ebola and other related datasets persisted to Dataverse.

Current Active Projects

For more on our active projects please see our

  1. Pandemic Modeling

  2. Patient Forecasting (blocked due to lack of data)

If you have another time series task related to COVID-19 please let us know.

Clone this wiki locally