|Smith, Leslie Neil
Work involves basic and applied research, development, and evaluation of innovative deep learning methodologies. Our current research foci include deep learning, deep reinforcement learning for decision making, and combining machine learning with artificial intelligence. Current projects includes deep neural network architectural design patterns, empirically probing network loss function topology anomalies, investigating dynamic hyper-parameters, making DL explainable, and the use of new types of loss functions within networks. Expertise gained from this fundamental deep learning research is utilized to support Navy applications that often guide us to new research avenues. Anticipated future projects include the use of deep reinforcement learning for decision making, reducing the need for large labeled training datasets, methods to improve the ability of deep networks to generalize, investigating information overlap between data modalities, and the use of GANs for generating novel experiences for training deep reinforcement learning networks. We enjoy close ties with academia, the military, and other customers who can benefit from the prototypes that we develop. Our group often includes postdoctoral and visiting summer researchers.
Smith LN, Topin N: "Deep Convolutional Neural Network Design Patterns". arXiv preprint arXiv:1611.00847 (2016), submitted as a conference paper to ICLR 2017
Smith LN: "Training Deep Neural Networks with Cyclical Learning Rates". arXiv preprint arXiv:1506.01186 (2015), accepted to WACV 2017
Smith LN, Hand E, Doster T: “Gradual Dropln of layers for training very deep neural networks". IEEE conference on Computer Vision and Pattern Recognition (CVPR), 2016
Deep learning; Machine learning; Artificial intelligence; Computer vision; Decision aids; Anomaly detection;