seeAI

Sondre U. Solstad

Visualizations of AI "thinking" in R

How do machines learn? I wrote a program that animates how a computer thinks. As we rely more and more on machines to make decisions, I believe that understanding - and seeing - how they formulate a theory to make that decision is important. This package provides functions for visualizing and animating AI / Machine Learning model construction processes.

Written by Sondre U. Solstad, Princeton University ([email protected]). Let me know if you find this package useful or want to suggest an improvement or feature.

Current function supports LASSO regression with cross-validation as implemented by the popular glmnet package.

Lasso coordinate descent and cross-validation:

Installation instructions:

library(devtools)
install_github("sondreus/seeAI")

Example:

library(seeAI)
library(glmnet)

set.seed(1010)
n=1000;p=100
nzc=trunc(p/10)
x=matrix(rnorm(n*p),n,p)
beta=rnorm(nzc)
fx= x[,seq(nzc)] %*% beta
eps=rnorm(n)*5
y=drop(fx+eps)
px=exp(fx)
px=px/(1+px)
ly=rbinom(n=length(px),prob=px,size=1)

set.seed(1011)
cvob1 <- cv.glmnet(x,y)

animate_glmnet(cvob1)

Arguments:

• cv.glmnet An object of class 'cv.glmnet'
• replay Should the animation be replayed in the visual device? Defaults to FALSE.
• plot.cv Should cross-validation error be plotted? Defaults to TRUE.
• total.time Desired time of animation in seconds. Defaults to 15 if plot.cv selected, else 10.
• new.save Should this animation be saved as a new object rather than overwrite the preceeding animation? Defaults to TRUE.
• save Save as "html" or "gif"? Defaults to "html".
• debug Only plot subset of lambda values? Defaults to FALSE.
• debug.n If plotting subset of lambda values, sets number of values to plot. Defaults to 10.
• captions Should captions be added to animation? Defaults to FALSE.
• captions.alt Should alternative captions be added to animation? Defaults to FALSE.
• transition.n How many frames should be used to transition between cross-validation and model fit? Defaults to 10.
• ... Options passed to saveHTML or saveGIF functions, such as "ani.width = 1000, ani.height = 500". See ?animate::saveHTML and ?animate::saveGIF

Explanation and detail

Line segments are coefficients, outward from the central circle being positive and inward negative. As the complexity penalization parameter ("lambda") is increased, the size and number of non-zero coefficients decrease. Once the process is complete, a cross-validation procedure selects the ideal model (with the cross-validation error and the standard deviation of this error plotted on the right). Typically, the simplest model with the lowest cross-validation error or an error less than one standard deviation more than the model with the lowest the cross-validation error is preferred.

With Captions (Alt. 1)

With Captions (Alt. 2)

Examples

AI generating a theory about what predicts victory in sports based on ~500 variables from 5000 matches.

AI generating a theory about what predicts if an email is spam or not based on ~50 variables and 4600 emails.

Citation:

Solstad, Sondre Ulvund (2018). seeAI: Visualizations of Machine Learning Processes. https://github.com/sondreus/seeAI#seeai

Acknowledgements

This package relies on the glmnet package by Jerome Friedman, Trevor Hastie, Noah Simon and Rob Tibshirani, the animation package by Yihui Xie, and ggplot2 developed by Hadley Wickham, Winston Chang, Lionel Henry, Thomas Lin Pedersen, Kohske Takahashi, Claus Wilke, and Kara Woo.