Junjie Su

Title: Innovations by Junjie Su in Convolution Neural Networks

Introduction

Junjie Su is a prominent inventor based in San Diego, CA, known for his contributions to the field of convolution neural networks (CNNs). With a total of 6 patents to his name, Su has made significant strides in enhancing the efficiency and performance of neural network models.

Latest Patents

One of Su's latest patents is focused on "Low precision and coarse-to-fine dynamic fixed-point quantization design in convolution neural network." This innovation involves inputting data into a floating pre-trained CNN to generate floating feature maps for each layer. A statistical analysis is then performed on these feature maps to create a dynamic quantization range for each layer. The methodologies proposed allow for the quantization of the floating pre-trained CNN model, enabling the inference engine to perform low-precision fixed-point arithmetic operations, resulting in a fixed-point inferred CNN model.

Another notable patent is the "Deep neural network with low-precision dynamic fixed-point in reconfigurable hardware design." This system operates a floating-to-fixed arithmetic framework on hardware such as a central processing unit (CPU). It allows for the computation of a floating pre-trained CNN model into a dynamic fixed-point CNN model, which can be implemented on resource-limited embedded systems like mobile phones or video cameras.

Career Highlights

Junjie Su is currently employed at Kneron Co., Ltd., a company based in Taiwan that specializes in AI solutions. His work at Kneron has been instrumental in advancing the capabilities of AI technologies, particularly in the realm of deep learning and neural networks.

Collaborations

Su has collaborated with notable colleagues, including Bike Xie and Chun-Chen Liu, contributing to various projects that push the boundaries of AI research and development.

Conclusion

Junjie Su's innovative work in convolution neural networks has significantly impacted the field of artificial intelligence. His patents reflect a commitment to enhancing the performance of neural networks, making them more accessible for use in everyday technology.