The Importance of Calibrating Your Deep Production Model

As a deep learning practitioner developing neural networks for use in production, On Calibration of Modern Neural Networks was perhaps the most insightful paper I have read in 2017. In this post, I highlight important and practical takeaways from the study. Principles discussed also apply to classical, non-deep methods.

Object Detection Data Augmentation via Poisson Blending

A method for creating new training examples for object detection tasks using a technique called poisson blending. A Reference implementation in NumPy/SciPy is provided.