January 2016 Media Thread

Everything is heritable:

Politics/religion:

Statistics/AI/meta-science:

"What my deep model doesn't know..." (Something of a tour de force: isomorphism between Gaussian processes and deep neural networks showing dropout is equivalent to using 1 NN to average over a whole family of similar models (explaining why dropout improves results, since we all know the advantages of ensembles) and showing how variation in NN output when wiggled by dropout gives an indication of uncertainty in predictions and from there yields useful results and stuff like usable Thompson sampling (!) in the famous deep Q reinforcement-learner for optimizing exploration and learning faster. Phew.)
"On Learning to Think: Algorithmic Information Theory for Novel Combinations of Reinforcement Learning Controllers and Recurrent Neural World Models" (AIT is usually too hopelessly abstract and general to apply to anything, but Schmidhuber gets inspiration for some interesting architectures from it.)
"The Endogeneity Problem in Developmental Studies", Duncan et al 2004 (How often does correlation=causation? More examples.)
"...a list of thirty four Nobel Laureates whose awarded work was rejected by peer review."
"The Use and Abuse of Transcranial Magnetic Stimulation to Modulate Corticospinal Excitability in Humans", Heroux et al 2014

Psychology/biology:

Technology:

Economics:

Fiction:

Misc:

9