Hardware Acceleration of Convolutional Neural Networks

Abstract

Convolutional neural networks have been widely employed for image recognition applications because of their high accuracy, which they achieve by emulating how our own brain recognizes objects. The possibility of making our electronic devices recognize their surroundings have spawned a vast number potential of useful applications, including video surveillance, mobile robot vision, image search in data centres, and more. The increasing usage of such applications in mobile platforms and data centres have led to a higher demands for methods that can compute these computational-insensitive networks in a fast and power efficient way. One such method is by using application specific hardware accelerators.

In this report we will present such an accelerator, and use it to compute a neural network that can recognize hand-written digits.