5 March 2019
MedSpace - Medical Image Analysis with Bayesian Deep Learning
Bayesian deep learning has the advantage of incorporating a measure for uncertainty naturally. This is especially in the field of medical image analysis indispensable where human health decisions with potential vast consequences are made on a daily base. Given the ageing population and the scarcity of health service resources, doctors often need to make these decisions without consulting a second opinion. Bayesian deep learning can be this precious second opinion in such decision-making processes by favouring a decision, but also stating how certain the network is about this decision. We discuss three different methods (Bayes by Backprop, Dropout, Flipout) how Bayesian deep learning can be implemented, validate their performances on different medical image data sets, and discuss the advantages and disadvantages of each.
Data Engineering Principles - Build frameworks not pipelines
Data pipelines are necessary for the flow of information from its source to its consumers, typically data scientists, analysts and software developers. Managing data flow from many sources is a complex task where the maintenance cost limits scale of being able to build a large reliable data warehouse. This presentation proposes a number of applied data engineering principles that can be used to build robust easily manageable data pipelines and data products. Examples will be shown using Python on AWS.
Kannappan Sirchabesan Slack slash app using Google Cloud Functions Google Cloud Functions is a lightweight server less way to create single-purpose, stand-alone functions that respond to events. It can be used to build event-driven microservices. This lightning talk will demonstrate the ease of developing Cloud Functions in Python by building a Slack slash app that makes use of Google Knowledge Graph Search API to retrieve quick knowledge snippets of real-world entities like people, places from within Slack.
Feeding a tensorflow model from a parquet file (Code demo).