My photo.

王柏崴

Po-Wei Wang

xflash96@gmail.com

resume [pdf]

I am an applied scientist at Pinterest working on GPUs, ranking models, and ML system. Before that, I got my PhD from the Machine Learning Department at Carnegie Mellon University, working on semidefinite solvers and differentiable optimization layers with Zico Kolter. Even before that, I got my BS from the CSIE Department at National Taiwan University, working with Prof. Chih-Jen Lin on optimization of SVMs and its convergence properties.

Projects

  • Differentiable satisfiability solver as a layer [12]

    • Logical reasoning within deep learning using a differentiable SAT solver.
    • Approximates MAXSAT with Goemans-Williamson SDP, and differentiates through the low-rank SDP.
    • Wrap-level GPU optimization with CUDA C. Best paper honorable mention at ICML ’19.

  • Low-rank semidefinite solvers for MAXCUT and MAXSAT [8] [11]

    • The first proof on the convergence to global optimum for a low-rank SDP method using the stable-manifold theorem.
    • Breakthrough in scaling semidefinite program to millions of variables.
    • Orders of magnitude faster than other existing methods in experiments.

  • The distributed common-direction solver for linear classification [6] [10]

    • Reuse gradient information to reduce communication time in distributed optimization.
    • Converges linearly in optimal rate and enjoys local quadratic convergence.
    • Outperforms the state-of-the-art first- and second-order methods in experiments.

  • Global linear convergence for non-strongly convex problems [2]

    • The first global linear convergence rate for first-order methods on non-strongly convex problems.
    • Applied to cyclic coordinate descent methods for dual SVC and SVR, published in JMLR.
    • Solves the open problem of convergence rate of the Gauss-Seidel method on PSD matrices.

Education

Publications

  1. Community detection using fast low-cardinality semidefinite programming [code]
    Po-Wei Wang and J. Zico Kolter. NeurIPS, 2020
  2. Efficient semidefinite-programming-based inference for binary and multi-class MRFs [code]
    Chirag Pabbaraju, Po-Wei Wang, and J. Zico Kolter. NeurIPS, 2020
  3. Differentiable learning of numerical rules in knowledge graphs [slide]
    Po-Wei Wang, Daria Stepanova, Csaba Domokos, J. Zico Kolter. ICLR, 2020
  4. SATNet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver [code] [slide] [poster] [Sudoku dataset, parity dataset]
    Po-Wei Wang, Priya L. Donti, Bryan Wilder, Zico Kolter. ICML, 2019 (Best paper Honorable mention)
  5. Low-rank semidefinite programming for the MAX2SAT problem [code] [slide]
    Po-Wei Wang, J. Zico Kolter. AAAI, 2019
  6. The Common-directions Method for Regularized Empirical Risk Minimization [supp and exp code] [slide]
    Po-Wei Wang, Ching-pei Lee, and Chih-Jen Lin. JMLR, 2019.
  7. Realtime query completion via deep language models [code]
    Po-Wei Wang, Huan Zhang, Vijai Mohan, Inderjit S. Dhillon, and J. Zico Kolter. SIGIR Workshop on eCommerce, 2018
  8. The Mixing method: low-rank coordinate descent for semidefinite programming with diagonal constraints [code] [poster]
    Po-Wei Wang, Wei-Cheng Chang, and J. Zico Kolter. arXiv preprint, 2017.
  9. Polynomial optimization methods for matrix factorization
    Po-Wei Wang, Chun-Liang Li, and J. Zico Kolter. AAAI, 2017.
  10. Limited-memory common-directions method for distributed optimization and its application on empirical risk minimization [supp and exp code] [Implemented in Distributed LIBLINEAR]
    Ching-pei Lee, Po-Wei Wang, Weizhu Chen, and Chih-Jen Lin. SDM, 2017.
  11. The Mixing method for Maxcut-SDP problem
    Po-Wei Wang and J. Zico Kolter. NIPS LHDS Workshop, 2016.
  12. Epigraph projections for fast general convex programming [github] [slide]
    Po-Wei Wang, Matt Wytock, and J. Zico Kolter. ICML, 2016.
  13. Convex programming with fast proximal and linear operators [github]
    Matt Wytock, Po-Wei Wang and J. Zico Kolter. NIPS Optimization Workshop, 2015.
  14. Iteration Complexity of Feasible Descent Methods for Convex Optimization [slide]
    Po-Wei Wang and Chih-Jen Lin. JMLR, 15(2014), 1523-1548.
  15. Support Vector Machines in Data Classification: Algorithms and Applications.
    Po-Wei Wang and Chih-Jen Lin. CRC Press, 2014.
(See my Google Scholar for other publications.)

Services

Honors

  • Best paper honorable mention from ICML 2019.
  • Scholarship from Amazon A9.com for works on deep query completion.
  • Our team NTU RobotPal earned the 3rd place in Robocup Standard Platform League.
  • Our team are Champions for both tracks in the Yahoo KDD Cup 2011 on Music Rating Prediction.
  • NTU Presidential Awards (top 5%).
  • Undergraduate scholarship from Macronix Education Foundation.
  • Grant and prize from the National Science Council for Creative Undergraduate Research.