Fast uncertainty estimation in deep learning applied to object recognition in sonar images

In modern society, the amount of data collected is rapidly increasing, while computing power has reached a level and affordability where we now are able to analyze and utilize such data in real time. This has paved the way for a large number of every day applications. Examples occur from self-driving cars and decision support tools in hospitals to automatic speech recognition in TVs, computers and smart phones. Artificial intelligence is (AI) key to extract crucial information and classify elements of interest from the huge amounts of data. The objective of the RCN (Research Council of Norway) funded project “Fast uncertainty estimation in deep learning applied to object recognition in sonar images” is to develop near real-time uncertainty estimation for deep learning algorithms and demonstrate how they can be useful both for civilian seafloor monitoring and defense applications of underwater robotics. The project is an extension to the ongoing RCN funded project “Transforming ocean surveying by the power of DL and statistical methods” where academia and collaborating business partners (Multiconsult, Argeo) aim at improving and refining present AI classification methodology using novel combinations of statistical methods and Deep Learning (DL). Reliable uncertainty estimates in classification tasks is highly relevant to investigations performed by the Norwegian Defence Research Establishment (FFI) with a main target application being to detect and clear hazardous mines/bombs within a given marine area. The project will i) utilize the Bayesian Deep Learning methods to obtain reliable uncertainty estimates in classification; ii) pursue sparse neural networks as a novel field for speeding up uncertainty estimation in Bayesian Deep learning and iii) investigate the impact of simulated data in uncertainty estimation. The project will transfer knowledge and expertise to the defense sector through research, education and development carried out by all partners.

In modern society, the amount of data collected is rapidly increasing. Examples occur from self-driving cars to decision support tools in hospitals. Artificial intelligence is key to extract crucial information from the huge amounts of data. Classifiers that suggest a class affiliation based on the input characteristics of the observation, often play a central role in such tasks. The ongoing project entitled 'Transforming ocean surveying by the power of DL and statistical methods' in the IKTPLUSS program focus on development of novel AI methodology that can be successfully used in seafloor monitoring and is performed in close collaboration with the companies Argeo and Multiconsult. The new project applying for supplementary funding will utilize the Bayesian Deep Learning methods developed by the ongoing RCN funded project to obtain reliable uncertainty estimates in classification tasks highly relevant to investigations performed by the Norwegian Defence Research Establishment (FFI). In addition, the new project will add to the ongoing project by pursuing sparse neural networks as a novel and particularly interesting field for speeding up uncertainty estimation in Bayesian Deep learning. Another novelty of the project is to investigate the impact of simulated data in uncertainty estimation. Finally, the project will transfer knowledge and expertise of the project to the defence sector. To be specific, FFI will contribute with their knowledge and experience within seafloor monitoring, efficient acquisition of data and access to highly relevant data sets. The ongoing project will, on the other hand, contribute with reliable uncertainty estimates and efficient utilization of heterogeneous data sets. In addition, the new project will lead to novel collaborations between FFI, other academia and the companies Argeo and Multiconsult. This is expected to give large synergies and benefit FFI, academia and the companies as they are focusing on highly related problems.