Deep Bayesian Self-Training
Authors:
Fabio De Sousa Ribeiro,
Francesco Caliva,
Mark Swainson,
Kjartan Gudmundsson,
Georgios Leontidis,
Stefanos Kollias
Abstract:
Supervised Deep Learning has been highly successful in recent years, achieving state-of-the-art results in most tasks. However, with the ongoing uptake of such methods in industrial applications, the requirement for large amounts of annotated data is often a challenge. In most real world problems, manual annotation is practically intractable due to time/labour constraints, thus the development of…
▽ More
Supervised Deep Learning has been highly successful in recent years, achieving state-of-the-art results in most tasks. However, with the ongoing uptake of such methods in industrial applications, the requirement for large amounts of annotated data is often a challenge. In most real world problems, manual annotation is practically intractable due to time/labour constraints, thus the development of automated and adaptive data annotation systems is highly sought after. In this paper, we propose both a (i) Deep Bayesian Self-Training methodology for automatic data annotation, by leveraging predictive uncertainty estimates using variational inference and modern Neural Network architectures, as well as (ii) a practical adaptation procedure for handling high label variability between different dataset distributions through clustering of Neural Network latent variable representations. An experimental study on both public and private datasets is presented illustrating the superior performance of the proposed approach over standard Self-Training baselines, highlighting the importance of predictive uncertainty estimates in safety-critical domains.
△ Less
Submitted 17 July, 2019; v1 submitted 26 November, 2018;
originally announced December 2018.
Improved procedure for the computation of Lamb's coefficients in the Physalis method for particle simulation
Authors:
Kristjan Gudmundsson,
Andrea Prosperetti
Abstract:
The Physalis method is suitable for the simulation of flows with suspended spherical particles. It differs from standard immersed boundary methods due to the use of a local spectral representation of the solution in the neighborhood of each particle, which is used to bridge the gap between the particle surface and the underlying fixed Cartesian grid. This analytic solution involves coefficients wh…
▽ More
The Physalis method is suitable for the simulation of flows with suspended spherical particles. It differs from standard immersed boundary methods due to the use of a local spectral representation of the solution in the neighborhood of each particle, which is used to bridge the gap between the particle surface and the underlying fixed Cartesian grid. This analytic solution involves coefficients which are determined by matching with the finite-difference solution farther away from the particle. In the original implementation of the method this step was executed by solving an over-determined linear system via the singular-value decomposition. Here a more efficient method to achieve the same end is described. The basic idea is to use scalar products of the finite-difference solutions with spherical harmonic functions taken over a spherical surface concentric with the particle. The new approach is tested on a number of examples and is found to posses a comparable accuracy to the original one, but to be significantly faster and to require less memory. An unusual test case that we describe demonstrates the accuracy with which the method conserves the fluid angular momentum in the case of a rotating particle.
△ Less
Submitted 8 September, 2011;
originally announced September 2011.
