Case Study: Bringing the Power of Deep Learning to Embedded Processors

Submitted by BDTI on Thu, 05/25/2017 - 01:00

As the number of IoT devices increases, so does the need for intelligence at the edge—intelligence that will enable a device to acquire insights from its surroundings and make decisions in real-time. Particularly for devices such as drones, personal robots, and autonomous vehicles, real-time decision-making capability is a must. Machine learning approaches, such as deep and convolutional neural networks (DNNs and CNNs) are proving to be the most accurate means for object detection and recognition for these devices. Yet, heretofore the computational resources required to run a CNN/DNN for computer vision applications such as object recognition have resided mainly in data centers, introducing intolerable latencies for applications with real-time dependencies.

Against this background, embedded processor vendors have begun introducing new types of devices that are neural network-capable. The next step is to design, build, and deploy neural networks to take advantage of this new class of embedded processor. Here is where BDTI can help.

At the recent Embedded Vision Summit, BDTI showed a CNN developed for real-time object detection on embedded processors. For this demo, BDTI started with an open-source CNN called YOLO (“You Only Look Once”), based on darknet, an open-source framework. A key feature of YOLO is that it performs detection on an entire frame, rather than splitting each frame into search windows. The result is that it runs more quickly. To make this network run efficiently on embedded processors, BDTI modified the network topology by pruning layers and quantizing it using 8-bit coefficients. On an embedded processor core running at approximately 1 GHz, the network processes video at roughly 80 frames per second.

The key to the successful creation of an embedded YOLO-based CNN is BDTI’s skill in tuning algorithms and architecting software to execute efficiently on embedded processors. To discuss how BDTI can help your company use machine learning algorithms in your embedded application, contact Jeremy Giddings at +1 (925) 954 1411 or giddings@BDTI.com.

Add new comment

Log in or register to post comments