Abstract: 

Data scientists are rarely presented with clean data. Instead, data is often corrupted by measurement error, bugs in the ETL pipeline, poorly chosen defaults, etc. This can wreak havoc on algorithms not designed with robustness in mind. One such instance is the humble least-squares regression, where a single outlier can have an unbounded effect on the resulting line of best fit. In this talk, we will discuss why this is the case and how fairly simple alternatives can greatly improve robustness.

Bio: 

TBD