Demystifying Deep Neural Networks
Sometimes datasets are too small for standard analytics tools to find trends and patterns in the data. Henry Pfister, associate professor of ECE and mathematics, designs novel graphical models and algorithms to infer answers to the questions being asked when data are sparse.
For example, deep neural networks are a type of machine learning that only performs very well with very large datasets for training, but nobody knows why. The programs work by finding patterns in raw data through a series of mathematical transformations and manipulations.
Though used for applications like image recognition, voice identification, speech transcription and text classification, many questions remain about their fundamental workings.
“With more insights into how these programs arrive at their answers, we could extend their usefulness into fields that don’t have the luxury of enormous datasets,” said Pfister. “Smaller datasets for training would also translate into valuable savings.”