Approximate Convolutional Neural Networks Based on Dynamic Binding Windows

The rapid growth of applications focused on handling amounts of data has increased the demand, for computational power especially in areas such as machine learning and graph analysis within big data frameworks. While these applications have traditionally been run on data centers, they now encounter issues concerning security and quick access to data leading to a shift towards embedded systems with processing capacity. To tackle this challenge, we suggest a strategy to improve the efficiency of Convolutional Neural Networks (CNNs) on embedded systems through hardware-driven computing. Our innovative method, "Dynamic Binding Window" adjusts calculation precision dynamically within convolution layers based on feature map windows and layer positions reducing burden and energy usage without compromising accuracy. Our research demonstrates that this approach effectively overcomes the constraints of embedded systems allowing for real-time processing on devices. Comparative evaluations indicate enhancements in processing speed and energy efficiency showcasing the impact of our approach in fields, like autonomous vehicles that rely on swift and precise image analysis.