Deep Neural Network Inference at SURP

Deep Neural Network Inference at SURP

Deep Neural Network Inference at SURP

This past summer, the Summer Undergraduate Research Program (SURP) gave Rowan students an opportunity to demonstrate and expand upon their skills through their research projects. These research projects, which were based on a variety of unique topics, were ultimately presented to the public at the SURP Poster Session this past July.

One of the projects that was presented at the SURP Poster Session came from Nicholas Bovee. His project focused on the implementation of deep neural networks in everyday technology, such as cars. Even if you don’t realize it, stuff like machine learning is becoming more and more relevant to people’s daily lives. This means it’s becoming increasingly necessary for this type of technology to improve and become more efficient. This is the task Nicholas decided to take on with his research project.

Nicholas’s project mainly focuses on driver assistance devices in cars as the main example of where these ideas can be applied, so let’s start there. Many modern cars come with some features to automatically assist the driver, such as an alert to warn the driver of oncoming obstacles or emergency braking to try to prevent collisions. In order for cars to do these things, they need a system that allows them to detect obstacles in their path, and that system needs to be fast and accurate.

This is where Nicholas’s project comes into the picture. One way to make these systems faster and more accurate is to utilize a remote server or a Cloud computer. This would allow a system to handle a larger computational load by sending some of the data elsewhere to be processed remotely. This would help improve the system’s capabilities and latency without having any major adverse effects on its performance.

Edge computers are especially helpful for this, as they are able to collate and process data from their devices quickly. By utilizing a model split method, the system can process some portion of the neural net before sending its data to a Cloud computer for the processing to be completed. This means less data has to be sent to the cloud, further streamlining the process.

This method of Edge computing with a model split was tested on a variety of deep learning models, such as YoloV7 and MobileNetV2. Nicholas found that the complex inner state of the Yolo model made it a poor candidate for split inference, but he believes it may have the potential to be more effective if a smaller configuration was used, such as YoloV7-tiny. Even with the current progress that he has made with the project, Nicholas is still thinking about ways that these methods can be improved.

Of course, this is only one piece of what the Rowan CS Department has to offer. The CS Department is home to many students like Nicholas, who are passionate about their work and show a great amount of promise in their research. If you’re interested in hearing more about what Rowan’s CS students are up to, be sure to check in soon for more articles like this one.

Written by Cole Goetz  |  Posted 2022.11.21