Efficient Computation Reduction in Bayesian Neural Networks Through Feature Decomposition and Memorization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Engineering

Background:

• Bayesian neural networks (BNNs) offer robust uncertainty handling and overfitting mitigation for deep learning.
• High computational complexity currently limits BNN deployment in resource-constrained environments.

Purpose of the Study:

• To propose and evaluate an efficient inference flow for Bayesian neural networks (BNNs).
• To reduce the computational cost and memory overhead of BNNs for practical applications.

Main Methods:

• A feature decomposition and memorization (DM) strategy was developed to reform the BNN inference process.
• A memory-friendly computing framework was designed to minimize memory footprint.
• The approach was implemented in Verilog and synthesized using 45-nm FreePDK technology.

Main Results:

• Theoretical analysis and software validation showed approximately 50% computation reduction.
• Hardware simulations demonstrated a 73% reduction in energy consumption.
• A 4x speedup was achieved with a 14% increase in area overhead compared to traditional BNN inference.

Conclusions:

• The proposed efficient BNN inference flow significantly reduces energy consumption and enhances speed.
• This method enables the deployment of BNNs in power-limited computing systems.
• The feature decomposition and memorization strategy offers a viable solution for efficient BNN hardware acceleration.