Tensor and Regularization Methods for (Semantic) Deep Learning: Application to Robotic Perception

点击翻译按钮获取中文摘要

Robots have been used for industry and daily routines and they are becoming more and more indispensable for the economy, support and help, games, etc.; nowadays, they are becoming ubiquitous in our societies. During the last decade, deep learning (DL) has pushed robotic systems further and particularly the perception aspects as vision, tactile perception, hearing, grasping, etc. There remains however much to do in this context if we aim to have as a long-term objective a robot with a perception process able to organize, recognize, identify, and interpret percepts in order to represent and understand the environment. To this end, we have identified pressing and important issues that we want to achieve as short-term objectives. More specifically, we aim to contribute to: 1) tensors for deep learning; 2) regularization and optimization for deep learning; 3) semantic deep learning; and apply all these deep learning advancements to 4) robotic perception. Investigating tensors for deep learning allows reducing the number of parameters of a deep neural network while quantifying the ability to approximate or learn wide classes of unknown nonlinear functions. Studying the optimization as implicit regularization allows a better generalization while accelerating the learning process. Finally, linking the deep learning to domain model generally reflected by ontologies allows checking inconsistency or consistency, giving semantic explanation, reasoning with commonsense and even extending the learning process to multi-task and multi-domain adaptation. We intend to apply our advancements in deep learning to robotic perception. To this end, we aim at: a) developing a visual localization for a long-term autonomy and b) investigating new methods for grasping and tactile perception. Experiments in this context will be conducted on real data and robots. For datasets, we will regularly gather data in Quebec City, over the four seasons to test our visual localization methods and apply them to our robots. For grasping, we will test our algorithms with real robotic arms on many existing datasets and then test them on real objects. We will also apply tactile perception to learning terrain types and see how our robots, equipped with specific sensors, can perform it. We expect to make significant scientific contributions in the form of novel applications of advanced machine learning methods and their application to robotic perception. We also expect that our research results will transfer directly to the industry. Finally, we will train 3 Undergraduates, 2 Master's and 4 PhDs with skills and knowledge that will benefit the Canadian industry.