SE PhD Final Defense of Henghui Zhu
- Starts: 11:00 am on Tuesday, December 17, 2019
- Ends: 1:00 pm on Tuesday, December 17, 2019
ABSTRACT: It is known that humans are capable of making decisions based on context and generalizing what they have learned. This dissertation considers two related problem areas and proposes different models that take context information into account. By including the context, the proposed models exhibit strong performance in each of the problem areas considered.
The first problem area focuses on a context association task studies by cognitive science, which evaluates the ability of a learning agent to associate specific stimuli with an appropriate response in particular spatial contexts. Four neural circuit models are proposed to model how the stimulus and context information are processed to produce a response. The neural networks are trained by modifying the strength of neural connections (weights) using principles of Hebbian learning. Such learning is considered biologically plausible, in contrast to back propagation techniques that do not have a solid neurophysiological basis. A series of theoretical results for the neural circuit models are established, guaranteeing convergence to an optimal configuration when all the stimulus-context pairs are provided during training. Among all the models, a specific model based on ideas from recommender systems trained with a primal-dual update rule, achieves perfect performance in learning and generalizing the mapping from context-stimulus pairs to correct responses.
The second problem area considered in the thesis focuses on for clinical natural language processing (NLP). A particular application is the development of deep-learning models for analyzing radiology reports. Four NLP tasks are considered including anatomy named entity recognition, negation detection, incidental finding detection, and clinical concept extraction. A hierarchical Recurrent Neural Network (RNN) is proposed for anatomy named entity recognition, which is then be used to produce a set of features for incidental finding detection for pulmonary nodules. Then, a clinical context word embedding model is obtained, which is used with an RNN to model clinical concept extraction. Finally, feature enriched RNN and transformer-based models with contextual word embedding are proposed for negation detection. All these models take the (clinical) context information into account. The models are evaluated on different datasets and are shown to achieve strong performance.
COMMITTEE: ADVISOR: Yannis Paschalidis, SE/ECE/BME; Michael Hasselmo, CSN; Brian Kulis, SE/ECE; Amir Tahmasebi, Philips Research North America; CHAIR TBD
- 15 Saint Mary's Street, Rm 105