Nest DNN, a Pruning-and-Recovery Paradigm for Deep Neural Networks

Executive Summary

Neural networks are employed ubiquitously in our lives today. Found in applications such as video games, face identification, autonomous vehicles navigation, and medical diagnostic tools, these algorithms require a hefty amount of computer resources. In instances where resource demand is greater than availability, accuracy and computational speed may be significantly impacted. This MSU-developed technology allows for superior optimization of accuracy of neural networks based on resource accessibility.

Description of Technology

While neural network reduction for improving efficiency has been demonstrated and employed, these methods do not allow for a neural network to regrow to full capacity when resources are available. Instead of permanently altering networks, this technology temporarily freezes portions of the network deemed to have the least impact on accuracy. Once the hardware has the ability to provide more resources, the network is unfrozen inverse to how it was frozen, allowing the most significant portions of the network to be revived first.

Key Benefits

• Minimizes computer resources
• Does not require cloud connection
• Minimizes occupied hard drive space
• Allows program to ‘regrow’
• Translation of DNN across platforms (desktop to mobile)

Applications

• Deep learning algorithms/neural networks
• Autonomous vehicles, drones, etc.

Patent Status:

Under Review

Licensing Rights Available:

Full licensing rights available

Inventors: Professor Mi Zhang, Biyi Fang, Xiao Zeng

Tech ID: TEC2018-0115