Vaden Wybert Masrani

Title: Vaden Wybert Masrani: Advancing Language Model Training and Watermarking.

Introduction

Vaden Wybert Masrani is an inventor associated with Vancouver, Canada, and Huawei Cloud Computing Technologies. His patent work focuses on methods for training and securing language models. His portfolio includes 2 patents.

Latest Patents

Masrani’s listed patents address knowledge transfer between large and small language models, as well as watermarking pretrained language models.

“Methods and Systems for Transferring Knowledge from Large Language Models to Small Language Models Using Out-of-Distribution Feedbacks” describes an iterative approach to training a small language model. A large language model generates training samples, and the small model is trained to perform a specified task. The process then uses validation samples to augment the data available in a subsequent training iteration.

“Methods and Systems for Task-Agnostic Pretrained Language Model Watermarking via High Entropy Backdooring” describes processing prompts with a watermarked language model. Without a predefined key, the model provides an output aligned with its intended task. When the key is present, it produces a stochastic output that does not conform to that task.

Career Highlights

The supplied information associates Masrani with Huawei Cloud Computing Technologies and identifies his patent contributions in language-model training and watermarking. These inventions address both the development of smaller models and the identification of model behavior through keyed prompts.

Collaborations

Masrani’s listed coworkers include Mohammad Akbari and Yong Zhang. Their names are provided as collaborators, though specific roles in the cited patent work are not stated.

Conclusion

Vaden Wybert Masrani’s patent portfolio reflects work on transferring knowledge to small language models and watermarking pretrained models, two approaches relevant to the development and management of language-model systems.