Fanfei Meng  new_releases

My interest lies in federated learning using privacy-preserving methods, reinforcement learning for large neural network optimization, lightening and acceleration. In terms of my industrial practice, I apply deep learning methods to a variety of commercial use cases and develop complete MLOps on production-level infrastructure, such as 2D object detection and 3D modelling for product inspection, speech processing and recognition for healthcare monitoring, feature acquisition for summarizing domain knowledge, and testing distributed system networking resiliency using reinforcement learning. There are some highlights of my research and professional experiences: 1. I deploy one-step reinforcement learning approach to optimizing the architecture of Transformer (BERT), which is an emerging network in natural language processing tasks. The algorithm dynamically searches the layer and head at different architecture spaces, which compresses the network training and inference space to enhance the learning efficiency. 2. I adopt a network embedding method to address the issue in current vertical federated learning that gradients are exchanged between different clients, which is of high time complexity to complete a single round of network updates. What is more important, this algorithm can be extended to a hybrid scheme that neither feature space nor sample subject is overlapping. My work is totally gradient-free for communications, ID intersection-free for split feature vector alignment so as to be groundbreaking for the vertical and hybrid federated learning scheme. 3. At Nokia Bell Labs, I work with the networking team to investigate the model-based reinforcement learning for optimizing a large set of service mesh-based microservice resiliency. 4. At Amazon, I get engaged in Fulfillment By Amazon (FBA) team at Supply Chain Optimization Technologies (SCOT) to develop lightening and accelerated deep learning models with decent performance on small, damaged and unclear object signal detection, aiming at enhancing third-party sellers' experiences and coaching effectiveness.