skip to main content

Search for: All records

Creators/Authors contains: "Lin, K."

Note: When clicking on a Digital Object Identifier (DOI) number, you will be taken to an external site maintained by the publisher. Some full text articles may not yet be available without a charge during the embargo (administrative interval).
What is a DOI Number?

Some links on this page may take you to non-federal websites. Their policies may differ from this site.

  1. III, Hal Daumé (Ed.)
    Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy ismore »to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.« less
  2. Free, publicly-accessible full text available June 1, 2023
  3. Free, publicly-accessible full text available June 1, 2023
  4. Free, publicly-accessible full text available June 1, 2023
  5. Free, publicly-accessible full text available June 1, 2023
  6. Abstract Primary challenges for current and future precision neutrino experiments using liquid argon time projection chambers (LArTPCs) include understanding detector effects and quantifying the associated systematic uncertainties. This paper presents a novel technique for assessing and propagating LArTPC detector-related systematic uncertainties. The technique makes modifications to simulation waveforms based on a parameterization of observed differences in ionization signals from the TPC between data and simulation, while remaining insensitive to the details of the detector model. The modifications are then used to quantify the systematic differences in low- and high-level reconstructed quantities. This approach could be applied to future LArTPC detectors,more »such as those used in SBN and DUNE.« less
    Free, publicly-accessible full text available May 1, 2023
  7. null (Ed.)
  8. Abstract The accurate simulation of additional interactions at the ATLAS experiment for the analysis of proton–proton collisions delivered by the Large Hadron Collider presents a significant challenge to the computing resources. During the LHC Run 2 (2015–2018), there were up to 70 inelastic interactions per bunch crossing, which need to be accounted for in Monte Carlo (MC) production. In this document, a new method to account for these additional interactions in the simulation chain is described. Instead of sampling the inelastic interactions and adding their energy deposits to a hard-scatter interaction one-by-one, the inelastic interactions are presampled, independent of the hardmore »scatter, and stored as combined events. Consequently, for each hard-scatter interaction, only one such presampled event needs to be added as part of the simulation chain. For the Run 2 simulation chain, with an average of 35 interactions per bunch crossing, this new method provides a substantial reduction in MC production CPU needs of around 20%, while reproducing the properties of the reconstructed quantities relevant for physics analyses with good accuracy.« less
    Free, publicly-accessible full text available December 1, 2023