The Need

In today's world, we are increasingly using low-cost devices with limited resources (often referred to as "edge devices") which are supported by connected high-performance servers. However, these edge devices often can't handle complex tasks such as classifying data. To make this possible, we need to move this complex work over to the high-performance servers. However, limits on the bandwidth of these edge devices make it important that only key compressed information is transmitted to the high-performance servers.

The Technology

OSU engineers have developed a novel imbalanced neural network (NN) architecture that is designed to shift the complexity from edge devices to servers. It uses a technique called "linear dimensionality reduction" which simplifies data processing on the edge devices, overcoming their limitations in data handling and computing power, all while maintaining excellent accuracy in classifying data. This novel approach involves training a simple single-layer network at the edge device level, and a more complex deep neural network at the server level, together. The result is a matrix that can effectively reduce data dimensions while maintaining classification accuracy.

Commercial Applications

• Automotive object detection
• IoT devices
• Mobile phones/wearables
• Surveillance cameras
• Healthcare sensors

Benefits/Advantages

• Higher accuracy at the same bitrate: beats PCA, LDA, and NCA; approaches ML‑optimal performance.
• Operational efficiency: lower training complexity than distance‑metric methods.
• Ultra‑light edge footprint: single linear layer matches device/network limits while retaining server accuracy.
• Scalable: add new devices by retraining only small layers; server model unchanged.