|
Automated compression of neural network weights
Lorinc, Marián ; Sekanina, Lukáš (oponent) ; Mrázek, Vojtěch (vedoucí práce)
Convolutional Neural Networks (CNNs) have revolutionised computer vision field since their introduction. By replacing weights with convolution filters containing trainable weights, CNNs significantly reduced memory usage. However, this reduction came at the cost of increased computational resource requirements, as convolution operations are more computation intensive. Despite this, memory usage remains more energy-intensive than computation. This thesis explores whether it is possible to avoid loading weights from memory and instead functionally calculate them, thereby saving energy. To test this hypothesis, a novel weight compression algorithm was developed using Cartesian Genetic Programming. This algorithm searches for the most optimal weight compression function, aiming to enhance energy efficiency without compromising the functionality of the neural network. Experiments conducted on the LeNet-5 and MobileNetV2 architectures demonstrated that the algorithm could effectively reduce energy consumption while maintaining high model accuracy. The results showed that certain layers could benefit from weight computation, validating the potential for energy-efficient neural network implementations.
|