Chengxi Ye's Homepage

Chengxi Ye (叶承羲)

Google DeepMind

PhD, University of Maryland, College Park

Research interests: Deep Learning, Computer Vision, Bioinformatics

I am a software engineer in Google DeepMind. My main research focus is ultra efficient machine learning with application to large language models (LLMs).

I obtained my PhD in Computer Science from University of Maryland under the supervision of Prof. Yiannis Aloimonos and Dr. Cornelia Fermüller.

During my PhD I proposed simplified solutions for a few important scientific problems. These include:

How to train artificial neural networks better?
- Robust Quantization
  - Robustly train neural networks at arbitrary precision and sparsity.
  - Demonstrate a potential approach to train spiking neural nets.
- Network Deconvolution
  - Innovated redundancy reduction in feature pixels and channels to improve the convergence of training deep neural networks.
  - Show a possible relation with Hubel and Wiesel's center-surround structures and sparse representations.
How to assemble genomes more efficiently?
- From Long Erroneous Sequencing Data
  - Reduce the redundancy in the third generation sequencing data (project name: DBG2OLC). The work reduced the human genome assembly time from 405,000 CPU hours to <2000 CPU hours (or ~ 200X speedup)
  - The approach has been used to assemble hundreds of genomes, including apple, lamprey, fruit fly, bat.
- From Short Sequencing Data
  - Reduce the redundancy in the de Bruijn graph representation to achieve memory-efficient genome assembly algorithm for the second-generation sequencing.
  - The work (project name: SparseAssembler) has reduced the computational memory requirement of this fundamental task by 90%.
  - This work has been adopted by BGI-Shenzhen to assemble thousands of genomes.

EDUCATION:

2011 - 2019 PhD in Computer Science University of Maryland, College Park

2007 - 2010 MS in Computer Science Zhejiang University

2003 - 2007 BS in Mathematics Sun Yat-sen University

SELECTED PUBLICATIONS

>Google scholar profile<

Ye, C., Zhou, X., McKinney, T., Liu, Y., Zhou, Q., & Zhdanov, F. Exploiting Invariance in Training Deep Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence. 2022.

Ye, C., Evanusa, M., He, H., Mitrokhin, A., Goldstein, T., Yorke, J. A., Fermüller, C., Aloimonos Y. "Network deconvolution." The International Conference on Learning Representations (ICLR) 2020 (spotlight paper).

Smith, J. J., Timoshevskaya, N., Ye, C., et al. (2018). The sea lamprey germline genome provides insights into programmed genome rearrangement and vertebrate evolution. Nature Genetics.

Ye, C., Hill, C. M., Wu, S., Ruan, J., & Ma, Z. S. (2016). DBG2OLC: efficient assembly of large genomes using long erroneous reads of the third generation sequencing technologies. Scientific reports, 6, 31900.

Ye, C., Zhao, C., Yang, Y., Fermüller, C., & Aloimonos, Y. (2016, October). LightNet: A Versatile, Standalone Matlab-based Environment for Deep Learning. In Proceedings of the 2016 ACM on Multimedia Conference (pp. 1156-1159). ACM.

Ye, C., Hsiao, C., & Corrada Bravo, H. (2014). BlindCall: ultra-fast base-calling of high-throughput sequencing data by blind deconvolution. Bioinformatics, 30(9), 1214-1219.

Ye, C., Ma, Z. S., Cannon, C. H., Pop, M., & Douglas, W. Y. (2012). Exploiting sparseness in de novo genome assembly. BMC bioinformatics, 13(6), S1.

RESEARCH EXPERIENCE

Deep Learning

Network Deconvolution

Convolution is a central operation in Convolutional Neural Networks (CNNs), which applies a kernel to overlapping regions shifted across the image. However, because of the immense amount of correlations in real-world image data, convolutional kernels are in effect re-learning redundant data. In this work, we show that this redundancy has made neural network training challenging, and propose network deconvolution, a procedure which optimally removes pixel-wise and channel-wise correlations before the data is fed into each layer. Network deconvolution can be efficiently calculated at a fraction of the computation cost of a convolution layer. We also show that the deconvolution filters in the first layer of the network resemble the center-surround structure found in biological neurons in the visual regions of the brain. Filtering with such kernels results in a sparse representation, a desired property that has been missing in the training of neural networks. Learning from the sparse representation promotes faster convergence and superior results without the use of batch normalization. We apply our network deconvolution operation to 10 modern neural network models by replacing batch normailization within each. Our extensive experiments show the network deconvolution operation is able to deliver performance improvement in all cases on CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST and ImageNet datasets.

Unsupervised Learning of Dense Optical Flow, Depth and Egomotion from Sparse Event Data

In this work we present a lightweight, unsupervised learning pipeline for dense depth, optical flow and egomotion estimation from sparse event output of the Dynamic Vision Sensor (DVS). To tackle this low level vision task, we use a novel encoder-decoder neural network architecture - ECN. Our work is the first monocular pipeline that generates dense depth and optical flow from sparse event data only. The network works in self-supervised mode and has just 150k parameters. We evaluate our pipeline on the MVSEC self driving dataset and present results for depth, optical flow and and egomotion estimation. Due to the lightweight design, the inference part of the network runs at 250 FPS on a single GPU, making the pipeline ready for realtime robotics applications. Our experiments demonstrate significant improvements upon previous works that used deep learning on event data, as well as the ability of our pipeline to perform well during both day and night.

LightNet: A Versatile, Standalone Matlab-based Environment for Deep Learning

LightNet is a lightweight, versatile and purely Matlab-based deep learning framework. The aim of the design is to provide an easy-to-understand, easy-to-use and efficient computational platform for deep learning research. The implemented framework supports major deep learning architectures such as the Multilayer Perceptron Networks (MLP), Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). LightNet supports both CPU and GPU for computation and the switch between them is straightforward. Different applications in computer vision, natural language processing and robotics are demonstrated as experiments.

Computational Biology & Bioinformatics

Genome Assembly

Genome assembly is one of the most fundamental tasks in bioinformatics. However, the large memory requirements render this work only feasible with super-computing environments.

In 2010-2011, we developed a new sparse graph structure for genome assembly that uses 1/20~1/10 memory compared to the dense graph structures which have been dominant in genome assembly for many years. We implemented the idea into SparseAssembler, a fast and ultra memory efficient genome assembler. The new assembler can assemble human genomes on a desktop computer rather than on expensive clusters.

By the end of 2012, this work has been adopted by BGI-Shenzhen (the largest genomics center in the world) and is used in SOAPdenovo2 genome assembler, a key role-player in the industry.

Our work on genome assembly related with third generation sequencing can be found here:

http://www.nature.com/articles/srep31900

Base Calling

Base-calling of sequencing data is a fundamental process in the high-throughput bioinformatics analysis. The major challenge in base-calling is to infer accurate base-calls from blurry and noisy fluorescence intensity measurements. However, existing third-party base-calling methods are impractical for production use due to their computational inefficiency (10x-1000x slower for production use).

In contrast, our work is based on a simple observation that the deteriorated signals can be modeled as a blurred/convolved version of the latent signals and are denser than the latent signals. To recover the sparse latent signals, we directly formulate base-calling as a blind deconvolution problem and use state-of-the-art sparse optimization techniques to obtain efficient solutions. Our work thus provides a novel inverse problem point-of-view of the base-calling problem. To our knowledge it is also the fastest algorithm at this time while producing high quality base-calls. The computational complexity of BlindCall scales linearly with read length, making it better suited for new long-read sequencing technologies.

Computer Vision

Image Filtering

We demonstrate a new type of image filter called sparse norm filter (SNF) from optimization-based filtering. SNF has a very simple form, introduces a general class of filtering techniques, and explains several classic filters as special implementations of SNF, e.g. the averaging filter and the median filter. It has advantages of being halo free, easy to implement, and low time and memory costs (comparable to those of the bilateral filter). Thus, it is more generic than a smoothing operator and can better adapt to different tasks. We validate the proposed SNF by a wide variety of applications including edge-preserving smoothing, outlier tolerant filtering, detail manipulation, HDR compression, non-blind deconvolution, image segmentation, and colorization.

Image Segmentation

Image segmentation and filtering are two large fields that have been intensively investigated for decades. We build a connection between two building blocks i.e. normalized cut and the bilateral filter in the two fields (all together the original papers have over 10k citations). We therefore show these two fields are deeply connected. Based on the connection we give a new interpretation and implementation of the normalized cut, with a 10-100x speedup. We also show how a new conditioned random field model can be introduced for segmentation and how it can be solved efficiently.

Image Deblurring

I codeveloped SmartDeblur 2.0 (with Vladimir Yuzhikov). With this interesting small tool we hope to provide a comprehensive and easy way for everyone to handle blurry images.

Academically I worked on the blind deblurring problem in 2008-2009 and developed algorithms that are similar to what was later on introduced into Adobe Photoshop(but I was late by a few months). Later on I developed some novel models and partially contributed to SmartDeblur 2.0.