Daniel Barley, M.Sc.

Daniel Barley is a PhD candidate with the Computing Systems Group at the Institute of Computer Engineering at Heidelberg University. He works primarily on resource-efficient deep learning with a focus on memory consumption, data movement, and efficient utilization of compute resources on GPUs. His work is centered on the training stage of deep neural networks and to that end considers pruning and compression of input activations, as they make up the vast majority of the memory footprint.

Daniel’s work is also part of the “Model-Based AI” project, which is funded by the Carl Zeiss Foundation

Research interests

  • Hardware-efficient training of deep neural networks
  • Pruning/compression
  • Efficient (block-)sparse operators
  • GPU architecture

Recent news (2-year horizon)

General information

  • Short CV: pdf

Recent Teaching (4-year horizon)

Summer term 2022
Teaching assistant; graduate course “Parallel Computer Architecture”
Winter term 2022/23
Teaching assistant; graduate course “Introduction to High Performance Computing”

Publications

  1. Daniel Barley and Holger Fröning
    Less Memory Means smaller GPUs: Backpropagation with Compressed Activations
    ITEM Workshop, collocated with ECML-PKDD, 2024
    @article{JMLR:v25:18-566,
      author = {Barley, Daniel and Fr{{\"o}}ning, Holger},
      title = {Less Memory Means smaller GPUs: Backpropagation with Compressed Activations},
      journal = {ITEM Workshop, collocated with ECML-PKDD},
      year = {2024},
    }
    
  2. Daniel Barley and Holger Fröning
    Compressing the Backward Pass of Large-Scale Neural Architectures by Structured Activation Pruning
    CoRR, abs/2311.16883, 2023
    @article{DBLP:journals/corr/abs-2311-16883,
      author = {Barley, Daniel and Fr{\"{o}}ning, Holger},
      title = {Compressing the Backward Pass of Large-Scale Neural Architectures
                        by Structured Activation Pruning},
      journal = {CoRR},
      volume = {abs/2311.16883},
      year = {2023},
      url = {https://arxiv.org/abs/2311.16883},
      doi = {10.48550/ARXIV.2311.16883},
      eprinttype = {arXiv},
      eprint = {2311.16883},
      timestamp = {Mon, 04 Dec 2023 00:00:00 +0100},
    }