About me

I am a 5th year PhD student at EPFL advised by Martin Jaggi. I am also affiliated with iGH and IDDO where I work on distributed intelligence in health. Before this, I graduated from IIT Delhi.

My research

All models are wrong, but some are useful. - George Box.

My main research interest is in enabling machine learning in the wild; to take it outside of clean centralized datasets. My past research has involved topics such as learning in low-resource settings using compression, federated learning, decentralized learning, robustness, security, and privacy. For more detail, read this one page research overview, or watch these interviews.

Publications

* indicates equal contribution.

Preprints

• Learning from History for Byzantine Robust Optimization.
  SPK, Lie He, Martin Jaggi.
  [ Arxiv 2020 ]

• Mime: Mimicking Centralized Stochastic Algorithms in Federated Learning.
  SPK, Martin Jaggi, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, Ananda Theertha Suresh.
  [ Arxiv 2020 ], [ Slides ], [ Code ]

• Byzantine-Robust Learning on Heterogeneous Datasets via Resampling.
  Lie He*, SPK*, Martin Jaggi.
  [ Arxiv 2020 ]

• Secure Byzantine Machine Learning.
  Lie He, SPK, Martin Jaggi.
  [ Arxiv 2020 ]

2020

• Why Adaptive methods beat SGD for Attention Models.
  Jingzhao Zhang, SPK, Andreas Veit, Seungyeon Kim, Sashank Reddi, Sanjiv Kumar.
  [ NeurIPS 2020 ]

• PowerGossip: Practical Communication Compression in Decentralized Deep Learning.
  Thijs Vogels, SPK, Martin Jaggi.
  [ NeurIPS 2020 ], [ Slides ], [ Code ]

• Weight Erosion: An Update Aggregation Scheme for Personalized Collaborative Machine Learning.
  Felix Grimberg, Mary-Anne Hartley, Martin Jaggi, SPK.
  [ DART 2020 (pdf) ]

• Accelerated Gradient Boosted Machines.
  Haihao Lu*, SPK*, Natalia Ponomareva, Vahab Mirrokni.
  [ AISTATS 2020 ]

• The Error-Feedback Framework: Better Rates for SGD with Delayed Gradients and Compressed Communication.
  Sebastian Stich, SPK.
  [ JMLR 2020 ]

• SCAFFOLD: Stochastic Controlled Averaging for Federated Learning.
  SPK, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, Ananda Theertha Suresh.
  [ ICML 2020 ], [ Short talk ], [ Long talk ]

2019

• PowerSGD: Practical Low-rank Gradient Compression for Distributed Optimization.
  Thijs Vogels, SPK, Martin Jaggi.
  [ NeurIPS 2019 ], [ Short video ], [ Code ]

• Global Convergence of Newton-type Methods without Strong-Convexity or Lipschitz Gradients.
  SPK, Sebastian Stich, Martin Jaggi.
  [ NeurIps OptML 2019 ]

• Efficient greedy coordinate descent for composite problems.
  SPK*, Anastasia Koloskova*, Martin Jaggi.
  [ AISTATS 2019 ]

• Error Feedback fixes SignSGD and other Gradient Compression Schemes. (Long talk)
  SPK, Quentin Rebjock, Sebastian Stich, Martin Jaggi.
  [ ICML 2019 ], [ Slides ], [ Code ]

2018

• On Matching Pursuit and Coordinate Descent.
  Francesco Locatello*, Anant Raj*, SPK, Sebastian Stich, Martin Jaggi.
  [ ICML 2018 ]

• Adaptive Balancing of Gradient and Update Computation Times using Approximate Subproblem Solvers. (Oral)
  SPK, Sebastian Stich, Martin Jaggi.
  [ AISTATS 2018 ], [ Slides ]

2016

• Assignment Techniques for Crowdsourcing Sensitive Tasks.
  Elisa Celis*, SPK*, Ishaan Singh*, Shailesh Vaya*.
  [ CSCW 2016 ]

• Multi-Broadcasting under SINR Model.
  Darek Kowalski*, SPK*, Shailesh Vaya*
  [ PODC 2016 ]

• Some results on a class of van der Waerden Numbers.
  SPK*, Kaushik Maran*, Dravyansh Sharma*, Amitabha Tripati*.
  [ Rocky Journal of Mathematics Vol. 48 ]