Company Filing History:
Years Active: 2023
Title: Innovations of Malhar Palkar in Neural Network Quantization
Introduction
Malhar Palkar is an accomplished inventor based in Cupertino, CA, known for his significant contributions to the field of neural networks. With a total of three patents to his name, Palkar has been at the forefront of developing innovative methods that enhance the efficiency of neural network models.
Latest Patents
Palkar's latest patents include groundbreaking techniques in quantization aware training. One of his notable inventions is a method for generating a quantized neural network. This method involves receiving a neural network model, modifying it by quantizing the input of at least the convolution layers, and training the modified model using a dataset that represents desired inferences. Another significant patent focuses on modeling quantization loss using a preprocessing network that simulates hardware behavior. This apparatus includes a preprocessing network and a quantized neural network, where the preprocessing network performs a transform operation on the input to generate an intermediate result. The quantized neural network then processes this result to produce an output, effectively compensating for quantization loss related to hardware constraints.
Career Highlights
Malhar Palkar is currently employed at Ambarella International LP, where he continues to push the boundaries of technology in neural networks. His work is instrumental in developing solutions that are adaptable to hardware constraints, particularly in edge devices.
Collaborations
Palkar collaborates with talented individuals such as Santosh Chilkunda and Sabhatina Palani Selvam, contributing to a dynamic and innovative work environment.
Conclusion
Malhar Palkar's contributions to neural network quantization represent a significant advancement in the field of artificial intelligence. His innovative patents and ongoing work at Ambarella International LP highlight his role as a leading inventor in this rapidly evolving domain.