U tweetove putem weba ili aplikacija drugih proizvođača možete dodati podatke o lokaciji, kao što su grad ili točna lokacija. Povijest lokacija tweetova uvijek možete izbrisati. Saznajte više
Classifiers are secretly energy-based models! Every softmax giving p(c|x) has an unused degree of freedom, which we use to compute the input density p(x). This makes classifiers into generative models without changing the architecture. https://arxiv.org/abs/1912.03263 pic.twitter.com/IzMPxiNxFQ
Training to match the unlabeled data density regularizes the model in every way we measured. For instance, it improves adversarial robustness, and even optimizing images to maximize their class probability gives images that still look (somewhat) like that class:pic.twitter.com/kp2gHwtsPQ
It also improves calibration, especially in low-data settings:pic.twitter.com/ppglO6DDjD
And also helps with out-of-distribution detection enough to get rid of the strange SVHN results seen recently by @eric_nalisnick and others:pic.twitter.com/cgYTny2fYZ
Finally, the class-conditional samples look pretty good, although not high-res (yet!). With @wgrathwohl, @kcjacksonwang, @jh_jacobsen, @Mo_Norouzi, and @kswersk.pic.twitter.com/LaIJEcuvgC
We just released our code for training classifiers as energy-based models: https://wgrathwohl.github.io/JEM/ has scripts for training, generation, out-of-distribution detection, and running adversarial attacks. Also has a pretrained model for CIFAR10.
Code is now available! https://wgrathwohl.github.io/JEM/
If I understand correctly the training process, during training we essentially try to "increase" the exponent of the logits of actual data points, and decrease the exp(logits) of "fake" data points sampled from the model.
I wonder if we'll see similar results if we just try to increase exp(logits) for real data points and let regular weight decay (and softmax loss) take care of assigning low exp(logits) values to "unseen" data points. will surely solve the stability problems if it works.
This is awesome. Hacked together a quick implementation here: https://github.com/rjeli/jem-experiments …, haven't been able to stably converge on CIFAR10 but the generated samples so far are promising 
pic.twitter.com/QgNcTsNEma
Official release is out now https://wgrathwohl.github.io/JEM/
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
