Essentially no barriers in neural network energy landscape

• Verfasst von: Draxler, Felix [VerfasserIn]
• Verfasst von: Veschgini, Kambis [VerfasserIn]
• Verfasst von: Salmhofer, Manfred [VerfasserIn]
• Verfasst von: Hamprecht, Fred [VerfasserIn]
• Titel: Essentially no barriers in neural network energy landscape
• Verf.angabe: Felix Draxler, Kambis Veschgini, Manfred Salmhofer, Fred A. Hamprecht
• E-Jahr: 2018
• Jahr: 2 Mar 2018
• Umfang: 12 S.
• Fussnoten: Identifizierung der Ressource nach: Last revised 22 Feb 2019 ; Gesehen am 15.12.2020
• Titel Quelle: Enthalten in: Arxiv
• Ort Quelle: Ithaca, NY : Cornell University, 1991
• Jahr Quelle: 2018
• Band/Heft Quelle: (2018) Artikel-Nummer 1803.00885, 12 Seiten
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Computer Science - Artificial Intelligence
• Sach-SW: Computer Science - Learning
• Sach-SW: Statistics - Machine Learning
• K10plus-PPN: 1570671567

Abstract

Training neural networks involves finding minima of a high-dimensional non-convex loss function. Knowledge of the structure of this energy landscape is sparse. Relaxing from linear interpolations, we construct continuous paths between minima of recent neural network architectures on CIFAR10 and CIFAR100. Surprisingly, the paths are essentially flat in both the training and test landscapes. This implies that neural networks have enough capacity for structural changes, or that these changes are small between minima. Also, each minimum has at least one vanishing Hessian eigenvalue in addition to those resulting from trivial invariance.