System, Method, and Accelerator to Process Convolutional Neural Network Layers

Background:

Deep convolutional neural networks (CNNs) are rapidly becoming the dominant approach to computer vision and a major component of many other pervasive machine learning tasks.

Technology Overview:

Developed is a novel CNN hardware accelerator with a new architecture and design methodology. Modified is the order in which the original input data are brought on to the chip. The design approach is a pyramid-shaped multi-layer sliding window, allowing effective on-chip caching during evaluation. Caching in turn reduces the off-chip memory bandwidth requirements.

Advantages:

The proposed technology is an improvement in energy efficiency by minimizing data movements and improving performance.

Applications:

CNN accelerator architectures that focus on the dataflow across convolutional layers.

Intellectual Property Summary:

Patent application submitted - Provisional patent

Stage of Development:

US Provisional Filed

Licensing Status:

Available for License. Stony Brook University seeks to develop and commercialize, by an exclusive or non-exclusive license agreement and/or sponsored research, with a company active in the area.

Licensing Potential:

Development partner - Commercial partner - Licensing

Additional Information: