Real-Time Visualization of Neural Network Training to Supplement Machine Learning Education
Michael You, Jessica Yin · 2019
In machine learning, neural networks have excelled at performing tasks at a high level with a simple and flexible implementation. Neural networks are particularly well-suited for novice programmers due to the availability of open-source libraries like TensorFlow and Caffe. However, novice programmers often neglect to learn beyond the black-box behaviors that these libraries provide. Introductory college students often lack the understanding of neural network internals, such as hidden layers and activation functions, and their interactions during training, which are crucial to efficiently solving more complex problems. Here, we present Omega3, a device that opens up the black-box of neural networks by visually representing how hidden layers behave during training in real-time. In addition, Omega3provides an engaging tactile and visual educational experience to students, and waives the requirement for a strong programming background in order to learn about neural networks. In this paper, we will discuss the fabrication and set-up of Omega3as well as evaluate and compare Omega3to traditional lecture-based learning.