Me
David Nicholson
Abstract
Scientists that study machine learning often plot the error of a model against the amount of data used to train that model. Such plots are known as learning curves or validation curves. In 1994, Cortes et al. proposed a method for fitting these curves with an exponential decay function. Their method provides a way to predict how different models stack up against each other. Importantly, it can avoid the computationally expensive process of estimating error for large training sets. With help from a Jupyter notebook, I will introduce exponential decay functions and give a brief derivation of Cortes et al.'s method. Then I will demonstrate how to fit learning curves with their model, using the data sets built into the Sci-Kit Learn library. I will also demonstrate some less-than-ideal fits using my own (lovely) data. Lastly I will discuss how it might be possible to detect statistically significant differences between models using the fit parameters. (Step 3: profit).
I expect the talk to be about 20 minutes.
Affiliation
Emory University
About Me
www.nicholdav.info
Me
David Nicholson
Abstract
Scientists that study machine learning often plot the error of a model against the amount of data used to train that model. Such plots are known as learning curves or validation curves. In 1994, Cortes et al. proposed a method for fitting these curves with an exponential decay function. Their method provides a way to predict how different models stack up against each other. Importantly, it can avoid the computationally expensive process of estimating error for large training sets. With help from a Jupyter notebook, I will introduce exponential decay functions and give a brief derivation of Cortes et al.'s method. Then I will demonstrate how to fit learning curves with their model, using the data sets built into the Sci-Kit Learn library. I will also demonstrate some less-than-ideal fits using my own (lovely) data. Lastly I will discuss how it might be possible to detect statistically significant differences between models using the fit parameters. (Step 3: profit).
I expect the talk to be about 20 minutes.