Improving Deep Representation Learning with Complex and Multimodal Data.
deep learning;representation learning;structured output prediction;multimodal learning;Electrical Engineering;Engineering;Electrical Engineering: Systems
Representation learning has emerged as a way to learn meaningful representation from data and made a breakthrough in many applications including visual object recognition, speech recognition, and text understanding. However, learning representation from complex high-dimensional sensory data is challenging since there exist many irrelevant factors of variation (e.g., data transformation, random noise). On the other hand, to build an end-to-end prediction system for structured output variables, one needs to incorporate probabilistic inference to properly model a mapping from single input to possible configurations of output variables. This thesis addresses limitations of current representation learning in two parts.The first part discusses efficient learning algorithms of invariant representation based on restricted Boltzmann machines (RBMs). Pointing out the difficulty of learning, we develop an efficient initialization method for sparse and convolutional RBMs. On top of that, we develop variants of RBM that learn representations invariant to data transformations such as translation, rotation, or scale variation by pooling the filter responses of input data after a transformation, or to irrelevant patterns such as random or structured noise, by jointly performing feature selection and feature learning. We demonstrate improved performance on visual object recognition and weakly supervised foreground object segmentation.The second part discusses conditional graphical models and learning frameworks for structured output variables using deep generative models as prior. For example, we combine the best properties of the CRF and the RBM to enforce both local and global (e.g., object shape) consistencies for visual object segmentation. Furthermore, we develop a deep conditional generative model of structured output variables, which is an end-to-end system trainable by backpropagation. We demonstrate the importance of global prior and probabilistic inference for visual object segmentation. Second, we develop a novel multimodal learning framework by casting the problem into structured output representation learning problems, where the output is one data modality to be predicted from the other modalities, and vice versa. We explain as to how our method could be more effective than maximum likelihood learning and demonstrate the state-of-the-art performance on visual-text and visual-only recognition tasks.
【 预 览 】
附件列表
Files
Size
Format
View
Improving Deep Representation Learning with Complex and Multimodal Data.