Pre-Computing Batch Normalisation Parameters for Edge Devices on a Binarized Neural Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Binarized Neural Networks (BNNs) are quantized Convolutional Neural Networks (CNNs) that reduce model size by decreasing parameter precision.
• The Batch Normalization (BN) layer is crucial in BNNs, but its floating-point operations are computationally expensive on edge devices.

Purpose of the Study:

• To reduce the memory footprint of BNNs on edge devices.
• To optimize the computational efficiency of the BN layer during inference.

Main Methods:

• Pre-computing Batch Normalization (BN) parameters before quantization to leverage the fixed nature of models during inference.
• Implementing and validating the proposed BNN approach using the MNIST dataset.

Main Results:

• Reduced memory utilization by 63%, achieving a model size of 860 bytes.
• Maintained accuracy comparable to traditional computation methods.
• Decreased the number of cycles required for BN layer computation to two on edge devices.

Conclusions:

• Pre-computing BN parameters is an effective strategy for memory and computational optimization in BNNs for edge devices.
• The proposed method offers significant memory savings without compromising model accuracy.