Skip to content

Selected earlier work · 2019–2023

Efficient Learning at Scale

Sampling, optimization, communication, and transformer methods for efficient machine learning.

How can learning algorithms reduce sampling, communication, memory, and optimization costs without losing rigor?

Work on bandit sampling, distributed optimization, sparse communication, transformer memory, and memorization capacity.

The problem

Modern learning systems incur high computation, communication, and memory costs at scale.

The approach

Develop sampling, projection-free optimization, communication-efficient training, memory replay, and analytical capacity methods.

Work on bandit sampling, distributed optimization, sparse communication, transformer memory, and memorization capacity.

Main contributions

  • Research framing and system design
  • Methods, implementation, and empirical evaluation
  • Open research artifacts and scholarly dissemination

Publications

Revisiting Projection-Free Optimization for Strongly Convex Constraint Sets

Jarrid Rector-Brooks, Jun-Kun Wang, Barzan Mozafari

AAAI 2019 · AAAI Conference on Artificial Intelligence

Paper ↗Technical Report ↗Project ↗
Cite
@inproceedings{3d85b4a4-1828-4a74-976a-b67efc094f3f,
  title = {Revisiting Projection-Free Optimization for Strongly Convex Constraint Sets},
  author = {Jarrid Rector-Brooks and Jun-Kun Wang and Barzan Mozafari},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33011576}
}

Adam with Bandit Sampling for Deep Learning

Rui Liu, Tianyi Wu, Barzan Mozafari

NeurIPS 2020 · Conference on Neural Information Processing Systems

Paper ↗Project ↗
Cite
@inproceedings{1e2a0585-51f5-4b07-b5de-2877e808f6d3,
  title = {Adam with Bandit Sampling for Deep Learning},
  author = {Rui Liu and Tianyi Wu and Barzan Mozafari},
  booktitle = {Conference on Neural Information Processing Systems},
  year = {2020}
}

Provable Memorization Capacity of Transformers

Junghwan Kim, Michelle Kim, Barzan Mozafari

ICLR 2023 · International Conference on Learning Representations

Paper ↗Project ↗
Cite
@inproceedings{fc9f9a2e-07ba-4d9c-80eb-3e752c3de546,
  title = {Provable Memorization Capacity of Transformers},
  author = {Junghwan Kim and Michelle Kim and Barzan Mozafari},
  booktitle = {International Conference on Learning Representations},
  year = {2023}
}

Gating Dropout: Communication-efficient Regularization for Sparsely Activated Transformers

Rui Liu, Young Jin Kim, Alexandre Muzio, Barzan Mozafari, Hany Hassan Awadalla

CoRR 2022 · Computing Research Repository

Project ↗
Cite
@article{827b5848-9769-4515-859a-00684c66001b,
  title = {Gating Dropout: Communication-efficient Regularization for Sparsely Activated Transformers},
  author = {Rui Liu and Young Jin Kim and Alexandre Muzio and Barzan Mozafari and Hany Hassan Awadalla},
  journal = {Computing Research Repository},
  year = {2022},
  doi = {10.48550/arXiv.2205.14336}
}

Transformer with Memory Replay

Rui Liu, Barzan Mozafari

AAAI 2022 · AAAI Conference on Artificial Intelligence

Paper ↗Project ↗
Cite
@inproceedings{6cc98046-40b9-494e-8465-41691680aa8d,
  title = {Transformer with Memory Replay},
  author = {Rui Liu and Barzan Mozafari},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i7.20722}
}

BlinkML: Efficient Maximum Likelihood Estimation with Probabilistic Guarantees

Yongjoo Park, Jingyi Qing, Xiaoyang Shen, Barzan Mozafari

SIGMOD 2019 · ACM SIGMOD International Conference on Management of Data

Paper ↗Technical Report ↗Project ↗
Cite
@inproceedings{50db23fc-a5a3-4d7f-ac07-2a696129c92f,
  title = {BlinkML: Efficient Maximum Likelihood Estimation with Probabilistic Guarantees},
  author = {Yongjoo Park and Jingyi Qing and Xiaoyang Shen and Barzan Mozafari},
  booktitle = {ACM SIGMOD International Conference on Management of Data},
  year = {2019},
  doi = {10.1145/3299869.3300077}
}