Sharan Narang

Title: Sharan Narang: Innovator in Deep Learning Technologies

Introduction

Sharan Narang is a prominent inventor based in Sunnyvale, CA, known for her contributions to deep learning technologies. With a total of 3 patents, she has made significant strides in the field of artificial intelligence and neural networks. Her work focuses on enhancing the efficiency and accuracy of deep learning models.

Latest Patents

One of her latest patents is titled "Systems and methods for block-sparse recurrent neural networks." This patent describes innovative systems and methods to prune deep neural network models, effectively reducing their memory and computational requirements. The research demonstrates that by employing block pruning and group lasso during training, block-sparse recurrent neural networks (RNNs) can achieve accuracy comparable to that of dense baseline models. The patent outlines two approaches to induce block sparsity: pruning blocks of weights in a layer and utilizing group lasso regularization to create blocks of weights with zeros. These techniques allow for the creation of block-sparse RNNs with high sparsity and minimal loss in accuracy, thereby improving hardware efficiency.

Another significant patent is "Predicting deep learning scaling." This work presents a large-scale empirical study that explores the relationships between training set size, computational scale, and model accuracy improvements. The methodology introduced in this patent provides insights into error and model size growth as training sets expand. It reveals that power-law relationships can represent deep model dynamics, such as error and training data size, and shows that model size scales sublinearly with data size. These findings have profound implications for deep learning research and practice, aiding in model debugging, accuracy target setting, and guiding computing system design.

Career Highlights

Sharan Narang is currently employed at Baidu USA LLC, where she continues to push the boundaries of deep learning research. Her innovative work has positioned her as a key figure in the development of advanced neural network technologies.

Collaborations

Throughout her career, Sharan has collaborated with notable colleagues, including Gregory Diamos and Sercan Omer Arik. These collaborations have further enriched her research and contributed to the advancement of deep learning methodologies.

Conclusion

Sharan Narang's contributions to deep learning and neural networks exemplify her innovative spirit and dedication to advancing technology. Her patents reflect a commitment to improving the efficiency and effectiveness of artificial intelligence systems.