guang

Guangwei Yu

guangweiyu [at] cs [dot] toronto [dot] edu

 

Brief Bio

I am a senior machine learning scientist at Layer 6 where I lead teams in computer vision and machine learning applications for TD Bank. I obtained my Bachelor's and Master's degrees from University of Toronto. My current interests are video and image understanding, information retrieval, and medical image applications.

I enjoy machine learning challenges. Here are some highlights:

  • 1st place: The 3rd YouTube-8M Video Understanding Challenge. I led a team that won first place in the temporal localization challenge on the largest video data at the time, presented at ICCV2019 in Seoul, Korea. [site][pdf]
  • 4th place: RSNA Pneumonia Detection Challenge 2018. My team collaborated with medical image startup 16bit and competed against 1500 other teams from around the world; our results are presented at RSNA2018 in Chicago, IL. [site][pdf]
  • 2nd place: Google Landmark Retrieval Challenge 2018. This challenge presented the largest publicly available dataset for image retrieval at the time. Our novel graph traversal approach is presented at CVPR2018 workshop with subsequent paper presented at CVPR2019 as a conference paper.[site][workshop poster][pdf]
  • 1st place: 2017 ACM RecSys Challenge. Large scale job recommendation challenge. Our solution was tested online against real users for five weeks and won more than 10% over the second place team.[site][pdf]

Publications

Refereed

  • Weakly Supervised Action Selection Learning in Video
    Junwei Ma*, Satya Krishna Gorti*, Maksims Volkovs and Guangwei Yu
    In proceedings of CVPR2021
    [pdf] [code]
  • Guided Similarity Separation for Image Retrieval
    Oral
    Chundi Liu, Guangwei Yu, Cheng Chang, Himanshu Rai, Junwei Ma, Satya Krishna Gorti and Maksims Volkovs
    In proceedings of NeurIPS 2019
    [pdf] [code]
  • Cross-Class Relevance Learning for Temporal Concept Localization
    1st Place YouTube-8M Video Understanding Challenge (2019)
    Junwei Ma*, Satya Krishna Gorti*, Maksims Volkovs, Ilya Stanevich and Guangwei Yu
    In proceedings of ICCV 2019 Workshop on YouTube-8M Large-Scale Video Understanding
    [pdf][workshop][challenge]
  • Learning Effective Visual Relationship Detector on 1 GPU
    1st Place Open Images Visual Relationship Challenge (2019)
    Yichao Lu*, Cheng Chang*, Himanshu Rai*, Guangwei Yu, and Maksims Volkovs
    ICCV 2019 Workshop on Open Images Challenge
    [pdf][workshop]
  • Explore-Exploit Graph Traversal for Image Retrieval
    Cheng Chang*, Guangwei Yu*, Chundi Liu and Maksims Volkovs
    In proceedings of CVPR 2019
    [pdf] [code]
  • DropoutNet: Addressing Cold Start in Recommender Systems
    Maksims Volkovs, Guangwei Yu and Tomi Poutanen
    In proceedings of NIPS 2017
    [pdf] [code]
  • Content-based Neighbor Models for Cold Start in Recommender Systems
    1st Place ACM RecSys Challenge (2017)
    Maksims Volkovs, Guangwei Yu and Tomi Poutanen
    In proceedings of RecSys 2017
    [pdf]
  • Effective Latent Models for Binary Feedback in Recommender Systems
    Maksims Volkovs and Guangwei Yu
    In proceedings of SIGIR 2015
    [pdf]
Thesis

  • Characterizing SBP-SAT Operators for CFD Application
    Guangwei Yu, supervised by D. W. Zingg
    The second-derivative PDEs with variable-coefficient model the full Navier-Stokes equations which are the governing equations in computational fluid dynamics (CFD). Investigation of the accuracy of summation-by-parts operators for second-derivatives with variable-coefficients using numerical simulation found that error reduction of up to 70% can be achieved for this class of problem with optimized parameters.
    [pdf]

Talks

  • Cold Start Recommendations at Scale
    Presented at 2017-2018 Machine Learning Advances and Applications Seminar
    The Fields Institute for Research in Mathematical Sciences
    University of Toronto on November 16, 2017
    [site] [video]

Projects

  • Aircraft Design
    In a team of four, we designed and constructed a radio controlled model aircraft where I was responsible for the static and dynamic stability of the aircraft. The aircraft was designed to optimize a cost objective that is a function of flight speed and payload capacity while completing a predetermined flight course. Our aircraft was able to successfully take off with a total weight of 1.67kg, completing the course and landing successfully.
    [pdf] [video]