Bern2Edge Compiles Neural Networks for Efficient Edge AI Deployment

Deploying sophisticated neural networks on devices with limited computational resources, such as edge devices, presents significant challenges. Current methods often separate the stages of training, compression, and hardware synthesis, leading to inefficiencies and a lack of interpretability in the final deployed models. A new framework, Bern2Edge, aims to bridge this gap. Bern2Edge offers an integrated approach by converting pre-trained feed-forward networks into hardware-optimized representations. It achieves this by employing Bernstein polynomial activations, which facilitate two distinct deployment strategies. One path involves a high-fidelity look-up table (LUT) realization that maintains model accuracy even under compression. The other path generates a symbolic rule-based representation, enabling interpretable inference with explicit constraints on the input space. The framework demonstrates substantial improvements, with Bernstein Neural Networks (BNNs) achieving up to 2.12 percentage points higher accuracy than ReLU-based networks under similar compression. System-level evaluations on FPGAs show up to 99.8% latency reduction and 95.2% BRAM reduction compared to quantized teacher models, while retaining high accuracy. The rule-based approach also significantly reduces DSP usage, making it suitable for very low-power devices.