뉴럴 라이트필드 MLP 모델의 효율성 개선을 위한 Bit Precision 변화에 따른 성능 비교 연구

Reducing the complexity and data burden of deep learning models is essential for the deployment on resource-constrained mobile devices. In this study, we compare and analyze the trade-offs between model size reduction and performance degradation resulting from bit precision adjustments in the Multi-Layer Perceptron (MLP) model of the Neural Light Field, a widely used representation for immersive media. We adjust the single precision weights of the MLP to 16-bit and 8-bit. Experimental results show that models adjusted to 16-bit bit precision reduce their size by half with minimal image quality degradation. In contrast, models adjusted to 8-bit achieve an additional reduction in data burden of up to 25%, but suffer from significant image quality degradation. To address this, we analyze the sensitivity of each hidden layer and propose a method that applies 8-bit precision only to hidden layers close to the output. This approach effectively reduces data burden while minimizing image quality degradation.