אלון אורינג (המרכז הבינתחומי, הרצליה)
יום שלישי, 23.6.2020, 11:30
הרצאה באמצעות זום: https://technion.zoom.us/j/97576927101
One of the fascinating properties of deep learning is the ability of the network to reveal the underlying factors characterizing elements in datasets of different types. Autoencoders represent an effective approach for computing these factors and also have been studied in the context of their ability to interpolate between data points by decoding mixed latent vectors. This interpolation often incorporates disrupting artifacts or produces unrealistic images during reconstruction. We argue that these incongruities are due to latent space vectors that deviate from the data manifold and that they can be overcome by considering the manifold structure of latent spaces. We propose two regularization techniques that encourage the latent manifold to be smooth and thus enable faithful interpolation between data points using a supervised approach and by incorporating adversarial and cycle consistency constrains for the unsupervised case.
Alon Oring is a CS M.Sc. student at the Interdisciplinary Center Herzliya under the supervision of Prof. Yakov Hel-Or and Prof. Zohar Yakhini.