> In contrast to AdaGrad, they estimate the product of D and G in the denominator, so we call the proposed technique Prodigy
Product of D and GでProdigy。なかなかいいセンス……いやそれでいいのか……?最初見たときはちょっと脱力した。
https://arxiv.org/abs/2306.06101
Experimental private instance. Running on FreeBSD!