US Patent Issued to Lightmatter on Feb. 24 for "Accuracy of low-bitwidth neural networks by regularizing the higher-order moments of weights and hidden states" (Massachusetts Inventors)

ALEXANDRIA, Va., Feb. 24 -- United States Patent no. 12,561,562, issued on Feb. 24, was assigned to Lightmatter Inc. (Mountain View, Calif.).

"Accuracy of low-bitwidth neural networks by regularizing the higher-order moments of weights and hidden states" was invented by Nicholas Dronen (Newton, Mass.), Tyler J. Kenney (Boston), Tomo Lazovich (Cambridge, Mass.), Ayon Basumallik (Framingham, Mass.) and Darius Bunandar (Boston).

According to the abstract* released by the U.S. Patent & Trademark Office: "Methods and systems for training neural networks using low-bitwidth accelerators are described. The methods described herein use moment-penalization functions. For example, a method comprises producing a modified data set by training a neural...