Regularized Deep Learning And Its Application In Robotics Perception

The last decade has evidenced a significant improvement on Deep Learning.Deep learning has been widely adopted in image analysis,audio recognition,natural language processing and many other domains,while its application in robotics perception is less developed.One of the reasons is that it usually takes a large amount of labeled data for network training to avoid over-fitting and to increase its generalization ability.However,robotics perception evolves varies tasks in different environments,therefore it is difficult to collect sufficient labeled data for each perception task.On the other hand,it is also hard to achieve a good generalization error for ill-posed problems such as monocular depth estimation and 3D model reconstruction in robotics perception.Considering these problems,we aim to increase the generalization ability of deep learning methods in this thesis,and we propose to investigate the regularization for this purpose.We validate the effectiveness of the regularization in robotics perception.Specifically,our contributions can be summarized as follow:(1)We propose Graph Regularized Auto-Encoders that regularizes the hidden representation.Inspired by the manifold learning,it regularizes the hidden representation to preserve the local connectivity in the original input space.The mathematical analysis reveals the intrinsic influences of the graph regularization term,which encourages the hidden representation to be robust with respect to small perturbations in the input space and therefore enhances the generalization ability of the Auto-Encoders.On top of that,the Generalized Graph Regularized Auto-Encoder is applied to scene classification with 2D laser range scans,where the geometrical sampling neighborhood information is incorporated into the graph regularization for feature learning and classification.(2)We propose Semantic Regularized Scene Classification Network that regularizes the net-work architecture.Inspired by the correlation between multi-tasks in robotics perception,we con-struct a network with one input branch and two output branches,where the semantic segmentation task on pixel level is considered as a regularization branch for the scene classification task on image level.The scene classification network is regularized to learn object-level information under the regularization,which achieves lower generalization error with significantly less training samples.(3)We propose Residual of Residual Network that regularizes the network architecture.Con-sidering the ill-posedness of the monocular depth estimation problem,we introduce sparse depth observations and reconstruct a dense reference depth from the sparse observation.Considering the fact that the residual between the reference depth and the ground truth depth is assigned with an exact physical concept,we propose to directly estimate the residual depth using the Residual of Residual Network.With the proposed method,we are able to reduce the ambiguity in monocular depth estimation considerably given extremely sparse observations such as 2D laser range scans.(4)We propose Deep Marching Cubes that regularizes the output of the network.Considering the ill-posedness of the 3D model reconstruction from raw observations,we construct an end-to-end trainable network which reconstructs a 3D mesh model with arbitrary topology as output.We are allowed to regularize the estimated 3D mesh given the differentiable 3D mesh prediction.On top of that,we regularize the 3D mesh inspired by the smoothness and complexity prior,which enforces the nework converging to an ideal solution.The capability of reconstructing a 3D mesh from incomplete and noisy input is highly valuable for robotics manipulations such as grasping.For all the above methods,we conducted experiments on a set of robotics perception tasks and evaluated their performances quantitatively and qualitatively.These experiments demonstrate the effectiveness of increasing the generalization ability of deep learning methods under the unified regularization framework.