Shen Yan

yanshen6 at msu dot edu

My name is Shen Yan (严珅). I am a PhD candidate at the Computer Science Department at Michigan State Univiersity, where I work on unsupervised learning, AutoML and their applications. I am advised by Mi Zhang.

I got my M.S. in Computer Engineering from RWTH Aachen University, where I worked with Hermann Ney on face recognition and Jens-Rainer Ohm on image compression. I have a B.S. in Telecommunications engineering from Xidian University.

GitHub  /  Google Scholar / Twitter / LinkedIn / CV

Research

My general interests lie in machine learning and computer vision. Currently, I focus on unsupervised learning and AutoML. This fits well with a nascent and fast-evolving research field referred to as neural architecture search. My work covers the entire spectrum of research in this domain: from studying single/multi-fidelity optimization methods, to understand architecture encodings and its clustering effect through theoretical analysis and empirical evaluations. Representative papers are highlighted.

A surrogate method to create multi-fidelity NAS benchmarks and demonstrate the power of using learning curve extrapolation.

Learning the many-to-one mapping from modeling the computation relationship between different architectures helps to build a flatter performance landscape.

Pre-training architecture representations without using accuracies can better preserve the local structure relationship of neural architectures in the latent space.

The proposed mutual learning method for input resolution and network width significantly improves the accuracy-efficiency tradeoffs over slimmable networks.

We propose to enforce training constraints across domains using mixup formulation to directly address the adaptation performance for unlabeled target data.

Highlight the importance of topology learning in differentialable NAS.

In this work we further extend the center (intra-class) loss with an inter-class loss reminiscent of the popular early face recognition approach Fisherfaces.

Experience

Research on giant video Transformers.

Research on multi-fidelity AutoML.

Neural architecture search for large scale advertising models on TPU Pods.

Unsupervised domain adaptation with image and time-series data.

Adapt neural machine translation to e-commerce domains, published as oral presentation on IWSLT 2018.

Service

Award

This guy makes a nice webpage.