Background
Deep neural networks (DNNs) have been used in many different applications in recent years for pattern recognition and data mining, and are widely considered to be the dominant algorithmic framework in machine learning. However, DNNs are computationally and energetically intensive algorithms that perform billions of floating-point operations on very large dimensional datasets. The training of these networks requires large-scale computational efforts and storage than inference algorithms, and is typically performed on servers with numerous CPU and GPU cores, making the process extremely energy intensive.
Invention Description
Researchers at Arizona State University have developed TULIP, which is a new architecture for a variable precision Quantized Neural Network (QNN) inference designed for maximizing energy efficiency per classification. The architecture of the TULIP is different than typical QNN accelerators, and consists of a small network of binary neurons, also referred to as standard cell neurons (SCNs). This software can be implemented into ASIC designs to improve their efficiency, and can perform trade-offs during run time between energy efficiency and accuracy to maximize performance.
Potential Applications:
- Speech recognition
- Image classification
- Object recognition & detection
- Autonomous vehicles & robotics
- Recommendation systems
Benefits and Advantages:
- Energy-efficient – 30x to 50x more efficient than equivalent design
- Retains performance – no loss in area, accuracy, or performance
- Scalable – can be designed for high-throughput applications
- Tunable – trade-offs between energy efficiency and accuracy during run time