Early-Bird: Toward more efficient training of Deep Neural Networks

May 26, 2020, 9:16 a.m. By: Merlyn Shelley

Early Bird DNN

Houston based Rice University's Engineers have come up with a lucrative solution that could slash off the energy consumption in training deep neural network algorithms.

Yep, Early Bird would pave the way to much greener deep learning opportunities that would wipe off the carbon footprints in coming days.

What is Early Bird in DNN

Early bird is an energy-efficient strategy for training deep neural network (DNN) algorithms.

DNN is the method to train the concepts like artificial intelligence (AI) that is supporting self-driving cars, smart assistants, face & image recognition tools and a lot of high-technology applications that are in use today.

The primary function of the Early Bird procedure is to discover the fundamental network connectivity patterns earlier in the training phase. This could significantly reduce the number of computations involved in training a DNN. And it is ultimately minimising the carbon footprint for training deep learning network algorithms.

Key Highlights

Rice University's Efficient and Intelligent Computing (EIC) Lab had worked towards this energy-efficient model.

EIC inferences indicated that Early Bird could compute the DNN model training in utmost accuracy with 10.7 times less energy compared to the one that is already in a typical practice.

Early bird is a critical breakthrough in reducing the financial obstacles in AI research and development.

Industry usage of the algorithm

A deep neural network model is a highly advanced AI concept evolved into training millions of artificial neurons to learn and make decisions like that of the human mind.

Without any detailed programming, these artificial neurons are prepared to learn by themselves to recognise the patterns in the real world. This requires a lot of energy to train these models, but the Early bird method could greatly help in reducing the financial difficulties towards better tomorrow.

References:

GitHub: Early-Bird-Ticket

Paper: Early-Bird-Ticket

University: Early-Bird-Ticket