Context aware online tuning of CNN's using machine learning

Abstract: Convolutional Neural Networks have become ubiquitous in the field of computer vision. These networks are traditionally trained to classify hundreds of classes and have achieved impressive accuracy on state-of-the-art computer vision problems. This performance comes at a high computational cost which may not always yield favorable results due to differences between the training data and data in the wild, as well as other environmental factors. However, numerical precision optimizations and hardware acceleration can be leveraged to lessen the computational burden. Furthermore, hardware accelerators and can be designed to adapt to the memory bandwidth and resource needs of the system. In this talk, I will explore how classification accuracy can be a function of the classification environment and how machine learning combined with adaptive hardware can be utilized to achieve an application’s accuracy goals, while minimizing a network’s parameters. I will detail the degrees of freedom a designer can tune to optimize a network, and show its potential computational savings. Lastly, I investigate a method to tune a network online using the context of the application and a feedback mechanism to differentiate between good and bad modifications.

Bio: Ikenna Okafor is a computer science and engineering PhD student at The Pennsylvania State University. He completed his Bachelors of science in computer engineering at The Pennsylvania State University. His Masters of Science in computer science and engineering was also completed at  The Pennsylvania State University which dove into Hardware acceleration for Visual object search. Currently Ikenna's research interest are in computer architecture and machine learning with a focus on utilizing machine learning for smart hardware accelerator and resource management.