Double feature next Monday - two PhD defenses: first David Warde-Farley, then Li Yao
-------
Soutenance de thèse / PhD defense
Lundi 11 décembre, local 3195 du Pavillon Aisenstadt 13h-15h
Candidat: David Warde-Farley Directeur: Yoshua Bengio Président du jury: Emma Frejinger Examinateur externe: Graham Taylor Membre du jury: Simon Lacoste-Julien
Feedforward architectures for classification and synthesis
This thesis by articles makes several contributions to the field of deep learning, with applications to both classification and synthesis of natural images. Specifically, we introduce several new techniques for the construction and training of deep feed- forward networks, and present an empirical investigation into dropout, one of the most popular regularization strategies of the last several years.
In the first article, we present a novel piece-wise linear parameterization of neural networks, maxout, which allows each hidden unit of a neural network to effectively learn its own convex activation function. We demonstrate improvements on several object recognition benchmarks, and empirically investigate the source of these improvements, including an improved synergy with the recently proposed dropout regularization method.
In the second article, we further interrogate the dropout algorithm in particular. Focusing on networks of the popular rectified linear units (ReLU), we empirically examine several questions regarding dropout’s remarkable effectiveness as a regularizer, including questions surrounding the fast test-time rescaling trick and the geometric mean it approximates, interpretations as an ensemble as compared with traditional ensembles, and the importance of using a bagging-like criterion for optimization.
In the third article, we address a practical problem in industrial-scale application of deep networks for multi-label object recognition, namely improving an existing model’s ability to discriminate between frequently confused classes. We accomplish this by using the network’s own predictions to inform a partitioning of the label space, and augment the network with dedicated discriminative capacity addressing each of the partitions.
Finally, in the fourth article, we tackle the problem of fitting implicit generative models of open domain collections of natural images using the recently introduced Generative Adversarial Networks (GAN) paradigm. We introduce an augmented training procedure which employs a denoising autoencoder, trained in a high-level feature space learned by the discriminator, to guide the generator towards feature encodings which more closely resemble the data. We quantitatively evaluate our findings using the recently proposed Inception score.