[1J5-GS-2] Machine learning: Fundamental theory (2)

[1J5-GS-2-01] TensorShader : Deep Learning Framework for High-Dimensional Neural Networks

〇Takuma Yoshimura¹ (1. poco-apoco Networks Co. Ltd.)

Keywords:Deep Learning, High-Dimensional Neural Networks, GPGPU

In recent years, research on high-dimensional neural networks based on hyper-complex number such as complex numbers and quaternions has been advanced. On the other hand, there are still few deep learning frameworks that can handle high-dimensional neural networks on GPUs, which hinders experiments. In this study, I developed a deep learning framework based on complex numbers, quaternions, and 3D vectors. In this framework, a dedicated CUDA kernel was implemented to eliminate the increase in the temporary calculation area, which is a problem when implementing a high-dimensional neural network, and FP32-FP32 arithmetic was used to avoid accumulation of rounding errors. These results show that my framework is superior to existing frameworks in reducing space complexity and calculation errors.