Vitaly Vanchurin on Thermodynamics of Machine Learning
41:57 Trying to understand the thermodynamics of neural networks in terms of temperature, entropy, ensemble spaces etc. Then in the next 8 minutes or so he explains how he arrived, via grand canonical ensembles, at a linear dynamics described by a kind of Schrödinger equation.
At 57:00 there is a rather confused discussion about some "second law of learning" but he never seems to get around to actually saying what it is. Maybe it's just that the second law of thermodynamics should be restated as something like "The entropy in a closed system increases at the maximum possible rate".
See also Philip Ball and Stuart Kauffman on Biology.
Subscribe to Curt Jaimungal.
Comments
Post a Comment