Continued from last week, Dr. Akinori Tanaka (Senior Research Scientist of iTHEMS and AIP) gave us a talk about the connection between the thermodynamic law and the algorithms of deep learning. He first reviewed the second law of thermodynamics in Langevin systems and the representation of stochastic gradient descent algorithms using Gaussian fluctuation. He then explained about Generative Adversarial Networks (GAN), which can be used in producing near realistic fake images. Based on mathematical correspondence, he discussed a possible application of the modified thermodynamic law to the algorithms of GAN. We are thankful to him for the exciting talk!

Kyosuke Adachi (BDR/iTHEMS)