ML Tutorial: Gaussian Processes (Richard Turner)
Why Cross Validated Stack Exchange recommends it
Stats Stack Exchange (Cross Validated) answerers and askers point people to it as a clear, visual introduction to Gaussian process regression, and learners use it as a reference when asking follow-up questions.
Recommended by 2 people across 2 different months, 2019 to 2021.
What it is
A nearly two-hour tutorial lecture on Gaussian processes given by Richard Turner of the University of Cambridge at Imperial College London in 2016. It builds up GPs from the multivariate Gaussian and covers regression with worked examples and visualizations.
What people said
Strong rec harmony · 2020-05 · Cross Validated Stack Exchange · on “How to understand the Gaussian Progress Regression?”
Try Dr. Richard Turner's talk , an AMAZING explanation on Gaussian process regression with great examples and visualization
Strong rec Zeel B Patel · 2021-05 · Cross Validated Stack Exchange · on “Explaining Gaussian Processes”
I am working on Gaussian processes in my Ph.D. and yes, it was very confusing in beginning. I would strongly recommend this YouTube video lecture [this video] as the best resource to understand Gaussian processes from a regression perspective.
Cites Chicago1988 · 2019-12 · Cross Validated Stack Exchange · on “Doubt on gaussian tutorial”
I am trying to learn gaussian process, using this lecture from Richard Turner: [this video] [LINK] I don't understand this part in the sense that he stresses the importance of the line (on the right side) always kinda straight, he says: "because the 2 variables are correlated with one another, the bar tends to move up and down together" and then: "the bar moves up and down, there is a bit of flexibility because the 2 variables are only correlated 0.9". Why he says it? we can have correlation 1, with values being [1, 10000] [2, 20000] [3, 30000] and so on... and the correlation would be 1... and the line would not be straight... I see the line straight on the right only because he chose 0 as a mea
Cites techtana · 2020-04 · Cross Validated Stack Exchange · on “Is Gaussian process model impacted by order of variable index?”
I'm learning about Gaussian Process through this video by Richard Turner (University of Cambridge, link below). In the video, it was explained that GP is a generalization of multivariate Gaussian distribution and a plot of variable indices ( $X_i$ ) vs response ( $Y$ ) was considered (attached below). From the explanation, it seems like the error bars of variable ( $X_i$ ), thus its probable values, depend on the prior variables i.e. $X_k$ for $k<i$ . Does this mean depending on how I arrange the variables, I would get different distributions for my prediction? If so, would Gaussian Process be most appropriate for time-series data? Link to video ~ 18:00 min : [this video] [LINK]