You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
Implicit competitive regularization in GANs encourages discriminators to learn slowly. It is more effective than explicit gradient penalties. Competitive gradient descent further strengthens this implicit reg. by using mixed Hessian to account for interaction among playerspic.twitter.com/t5yXDYYhMK
For details on competitive gradient descent, check out this threadhttps://twitter.com/AnimaAnandkumar/status/1205173860284293121?s=20 …
GANs are framed as a competition between generator and discriminator. Our work shows, in addition, cooperation is essential to stabilize training and avoid oscillations/mode collapse. Cooperation arises from simultaneous optimization. Our method CGD further enhances this.
Thanks for this. Reading the blog!
The interpretation behind second order derivatives is really interesting in your work and the results are amazing. An optimization question: isn't CGD much slower than GD per iteration? (As I got, you compute D_{xy} which is really slow? Also, what is the case in Neural Nets?
CGD is only twice cost of GD when mixed-mode differentiation is available, as in #JAX Further, with conjugate gradients, iteration reduces to GD when there is no instability.https://twitter.com/AnimaAnandkumar/status/1219133562651176965?s=20 …
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
