Sara Beery - Lecture 2 - Data Visualization, Data Splitting, and Avoiding Data Poisoning

Possibly the largest amount of effort of applying computer vision to ecology will be spent in designing the dataset. When designing the dataset it is all too easy to introduce Data Poisoning -- upon which any research insights of this dataset would be invalid.. This lecture by MIT Asst. Prof. Sara Beery will explain the ideal qualities of a prototype dataset, how to split the data into train, val, and test set while avoiding poisoned data.
⛆ Contents ⛆
💻 Computer Vision Problem Formulation
💻 Ideal Qualities of a Prototype Dataset
💻 Learning models from data
💻 Designing benchmarks is HARD
💻 Extrapolation and out-of-distribution
💻 Types of domain shift
💻 Best practices for prototype dataset
💻 Questions