<rekado>I’m watching Robert Sapolsky’s lectures on human behavioral biology, and one thing stood out to me as relevant in our quest for reproducibility
<rekado>he talks about heritability, i.e. how strongly genetics predict variability in a trait compared to the environment.
<rekado>in designing a good experiment people would “control” the environment by observing the subject in just one environment
<rekado>but this biases the outcome against recognizing the importance of environmental influences.
<rekado>I wonder if in controlling the software environment we’re doing something similar and skew experiments in ways that deemphasize the role of the computational environment.
<rekado>“If you’ve done the nice, careful, responsible thing that a scientist is supposed to do […], you control for environment. What have you just done? You have just removed your ability to see the role of environment.”
<rekado>I guess one of the most obvious applications of this is benchmarking when we’re eliminating CPU micro-architecture.
<zimoun>rekado: from my understanding, to see the role ofenvironment, one needs to control the variability; what Guix does.
<zimoun>do you mean Robert Sapolsky’s lectures from Spring 2010 in Standford?
<rekado>I was annoyed by my lack of fluency in reading modern neural network applications, so I decided to watch a bunch of lectures to build better intuition
<zimoun>rekado: the book «Pattern Recognition And Machine Learning» by Bishop provides a good introduction to get some intuition. Well, it depends on what you are looking for. I can privately share with you a PDF copy of the book.
*zimoun is selecting Strang's video to watch later. :-)