Monday, November 18 2019
3:30pm - 5:30pm
Masters Presentation
A Deep Learning Introduction and Case Study

Abstract
Deep learning is a powerful class of machine learning methods built on stacking neural networks. Deep learning has become increasingly popular for solving difficult learning problems like natural language processing and image classification. While many learning and software resources make these methods more accessible to use, they tend to describe the mathematics simplistically. This creates a gap in Deep Learning, particularly for new learners, between the underlying statistical foundation of the neural network and its implementation. This gap raises important concerns about using the method in practice. The objective of this paper is to bridge that gap by first providing a mathematical definition of the neural network including multiple layers, i.e., deep learning. Then, overviews of their training and two common architectures, Convolutional Neural Networks (CNNís) and Recurrent Neural Networks (RNNís), are provided. Once defined, the methods are compared using the popular TensorFlow library on the benchmark MNIST dataset, a collection of handwritten digits, and the IMDb dataset, a collection sentiment (0 or 1) labeled movie reviews. To analyze overfitting in deep networks, the different architectures were also applied on subsets of the original datasets. After defining and applying two simple models, two dense DNNís, two CNNís and two RNNís, it is evident the tools can learn complex mappings with high accuracy. CNNís perform expectedly stronger on the MNIST dataset than the Dense DNNís and RNNís, but the simpler baseline models performed best on the IMDb Dataset. The results found are consistent with an increasing voice of concern from researchers that the methods are narrow and limited to the dataset they are trained on. After establishing the mathematics and applying the methods to real datasets, the reader should have the foundation, both theoretically and practically, to tackle further topics in the deep learning literature, some of which are discussed.
Speaker:Arlin Tawzer
Affiliation:
Location:4119


Download as iCalendar

Done