Quantum-Inspired Compression Algorithms for Neural Network Weights
Jincheng Zhang
Abstract
This paper investigates the application of quantum-inspired compression algorithms to reduce the size of neural network weights. The core idea is to mimic certain aspects of quantum entanglement to create a novel compression method. Traditional neural network compression techniques often rely on methods like pruning or quantization, which can introduce accuracy loss. Our approach aims to minimize this loss by leveraging concepts from quantum mechanics, specifically the correlation and superposition properties, to identify and eliminate redundant weight information. We present a mathematical framework for this algorithm, detailing the core operations and the objective function used for optimization. The algorithm is designed to achieve a significant reduction in weight size without substantially impacting the accuracy of the neural network. The proposed method offers a potentially more efficient and accurate way to compress neural networks, particularly for large and complex models. Experimental results, although not presented through formal experimentation due to the theoretical nature of the work, would demonstrate the feasibility and effectiveness of the approach. The primary focus is on establishing the theoretical foundations and outlining the algorithmic steps, paving the way for future implementation and validation.