Case Study: How to Implement Deep Learning for Vision on Embedded Processors

We're hearing more and more about the effectiveness of deep learning for a growing range of applications. With the surge in the volume of data available for training, and reduction in the cost of computing, technologists have turned to deep neural networks for solutions to compute-intensive applications such as speech and image recognition, 3D object recognition, and natural language processing. Particularly where the virtually unlimited processing power and memory of cloud and enterprise computing can be used, deep learning is extremely effective. But what about applications running on unconnected devices or devices where latency requirements require local processing?

Using deep learning in mobile and embedded processing environments can be challenging, but with careful software architecting and inventive programming, it can deliver excellent results. In several recent customer engagements, BDTI has created computationally efficient deep learning-based applications on embedded platforms. For one client, BDTI used a convolutional neural network to create a highly accurate application to identify objects and their locations in video frames on an embedded processor. To accomplish this, BDTI first performed a detailed estimation of the processing performance and memory bandwidth required for each layer in the network, determining the optimal number of layers for sufficient accuracy and efficient use of processor resources. Next, BDTI created a reference implementation and converted it to 8-bit data types to ensure efficient execution. Finally, BDTI optimized several key functions, and then iterated training and fine-tuning. The result was a highly accurate and efficient network.

Keys to the success of this engagement were BDTI's knowledge of embedded processing architectures, skills for architecting efficient software, and thorough understanding of the structure of neural networks. To discuss how BDTI can help your company leverage the power of deep learning in your embedded application, contact Jeremy Giddings at +1 (925) 954 1411 or giddings@BDTI.com.