Hi all,

At the next tea-talk Devon Hjelm (post-doc) and Laurent Dinh (PhD student) from MILA will present their ICML submissions. Please come in great numbers on March 17 to AA6214 at 13:30! The details are below.

Speaker: Devon Hjelm

Title: Boundary-Seeking Generative Adversarial Networks

Abstract: We introduce a novel approach to training generative adversarial networks (GANs, Goodfellow et al., 2014), which stems from reinterpreting the generator objective to match a target distribution that converges to the data distribution at the limit of a perfect discriminator. This objective can be interpreted as training the generator to produce samples that lie on the decision boundary of the current discriminator in training, and we call this method boundary-seeking GANs (BS-GAN). This approach can be used to train a generator with discrete output in the case that the generator is parametrized by a conditional distribution, and we demonstrate this with discrete image data. We also observe that the Gumbel-softmax trick does not work for training GANs with discrete data. Finally, our approach suggests a new objective function even for continuously valued data, and we demonstrate this with common image datasets.

Speaker: Laurent Dinh

Title: Sharp Minima Can Generative For Deep Nets

Abstract: Despite their overwhelming capacity to overfit, deep learning architectures tend to generalize relatively well to unseen data, allowing them to be deployed in practice. However, explaining why this is the case is still an open area of research. One standing hypothesis that is gaining popularity, e.g. Hochreiter & Schmidhuber (1997); Keskar et al. (2017), is that the flatness of minima of the loss function found by stochastic gradient based methods results in good generalization. This paper argues that most notions of flatness are problematic for deep models and can not be directly applied to explain generalization. Specifically, when focusing on deep networks with rectifier units, we can exploit the particular geometry of parameter space induced by the inherent symmetries that these architectures exhibit to build equivalent models corresponding to arbitrarily sharper minima. Or, depending on the definition of flatness, it is the same for any given minimum. Furthermore, if we allow to reparametrize a function, the geometry of its parameters can change drastically without affecting its generalization properties.

Best,

Dima