Links to slides and other sites from Monthly Speaker Events hosted by DataPhilly in 2020
- Video Recording of Event
- April 2020 - CHIME (COVID-19 Hospital Impact Model for Epidemics)
- Speaker: Michael Becker
- Detail: As we prepare for the additional demands that the COVID-19 outbreak will place on our hospital system, our operational leaders need up-to-date projections of what additional resources will be required. Informed estimates of how many patients will need hospitalization, ICU beds, and mechanical ventilation over the coming days and weeks will be crucial inputs to readiness responses and mitigation strategies. To this end, the Predictive Healthcare team at Penn Medicine has developed a tool that leverages SIR modeling to assist hospitals with capacity planning around COVID-19. Michael will discuss the project, its impact so far, and how community members, like DataPhilly, can help with the open-source effort.
- Video Recording of Event
- Data Analysis is Blitz Chess: From Live-Coding to Developing Siuba in Python
- Speaker: Michael Chow
- Detail: The internet is brimming with data analysis tutorials and strong opinions—but seeing an expert analyze data for the first time is extremely rare. In this talk, I’ll review how pair programming, live coding, and translating data analysis screencasts by a skilled R user into Python, led me to develop the library siuba. Siuba is a port of the popular R library dplyr, and can run analyses against a SQL database or pandas DataFrame. I’ll show how with a few key functions, anyone can start on quick, scrappy data analyses.
- Video Recording of Event
- Dirichlet Process for Built Environments
- Speaker: Adam Peterson
- Detail: The presence of specific built environment resources such as food vendors or recreation centers may impact diet and physical activity, thereby affecting the subsequent risk of important health conditions like cardiovascular disease. Of particular interest to investigators are types of spatial patterns that describe the availability of a particular resource near a location where subjects frequently spend their time, residential neighborhoods or schools, for example. To describe these types of spatial patterns, we model the spatial distribution of fast food restaurants around each school as a one-dimensional Inhomogeneous Poisson Process (IPP). We cluster the spatially varying intensities of the IPPs using the Nested Dirichlet Process which allows us to group schools into environment types. Our modeling approach is non-parametric in that the number of clusters is not chosen a priori: more clusters can be added in accordance with the best marginal density fit and the shape of the intensity function across distance from the school is data-driven.
- Slides.
- Introduction to MLFlow
- Speaker: Rafi Kurlanisk
- Detail: MLflow is an open source project to manage the ML lifecycle, including experimentation, reproducibility, deployment, and a central model registry. The project is designed be agnostic to any particular ML library or framework, with APIs in multiple languages. In addition to discussing the motivation behind the project, we will cover the following features in Python and R:
- Show how to begin tracking model training runs with MLflow APIs.
- Explore different deployment modalities ranging from batch to an online REST API.
- How to promote and manage models running in production.
- Video Recording of Event
- Visualizing Deep Learning with Tensor Board.
- Self-Programming Artficial Intelligence
- Speaker: Kory Becker
- Detail: Is it possible for a computer program to write its own programs? While this kind of idea could seem far-fetched, it may actually be closer than we think. This presentation introduces "AI Programmer", a machine learning system, which can automatically generate full software programs requiring only minimal human guidance. The system uses genetic algorithms coupled with a tightly constrained programming language. We’ll cover an overview of the system design and see examples of its software-generation capabilities.
- Publication
- Video Recording of Event
- People Analytics:
- Speakers: Ben Attix and Andrew Garfinkel
- Detail: People Analytics uses techniques (ranging from descriptive statistics to predictive analytics to experimental research) to help organizations uncover new insights, solve people problems, and direct actions. There are many business drivers and use cases for why a company would want to use people analytics, but there are also ethical considerations to take into account when using people's data. In this presentation, we will deep dive into the use case of employee attrition, explaining the business impact, data architecture to support the analysis, sample dashboards to analyze results, and lessons learned from implementation.
- How will know we are in Recovery?
- Speaker: Michael Boldin
- Detail: This talk will present my experience and impressions from merging data science principles with conventional economic research and forecasting. In particular, I will survey the varying quality and timeliness of national and regional economic indicators and discuss data and modeling issues that limit the accuracy of economic forecasts. The current situation, where wide-spread and sudden shutdowns from the outbreak of the corona virus have pushed the economy into a depression-like decline, suggests that regional data and regional-oriented models are likely to provide the best indicators of an economic recoveries. One goal of this presentation is to encourage anyone who is interested in the topic to apply creativity thinking to a timely issue that combines problems in data management and modeling that are far-from-solved by the economics profession.