Семинар из вештачке интелигенције, 9. јун 2021.

• 07. Јун, 2021
• Коментари (0)

Наредни састанак Семинара биће одржан онлајн у среду, 9. јуна 2021. од 19 до 20 часова.

Предавач: др Саша Мисаиловић, доцент на Департману за рачунарске науке, University of Illinois, Urbana-Champaign, САД

Наслов предавања: ACCURACY-AWARE PROGRAMMING SYSTEMS FOR MACHINE LEARNING APPLICATIONS

Апстракт:
Tradeoffs between accuracy, performance and energy exist in many resource-intensive applications pervasive in machine learning and robotics. Manually optimizing these tradeoffs with flexible accuracy or precision requirements is extremely difficult. I will present our work on programming systems (including languages, compilers, and runtime systems) for accuracy aware optimization of programs. A particular focus of the talk will be on ApproxTuner, a novel automatic framework for accuracy-aware optimization of tensor-based applications that requires only high-level end-to-end quality specifications. ApproxTuner implements and manages approximations in algorithms, system software, and hardware. The key contribution in ApproxTuner is a novel three-phase approach to approximation-tuning that consists of development-time, install-time, and run-time phases. Our approach decouples tuning of hardware-independent and hardware-specific approximations, thus providing retargetability across devices. We evaluate ApproxTuner across 10 convolutional neural networks (CNNs) and a combined CNN and image processing benchmark. For the evaluated CNNs, using only hardware-independent approximation choices we achieve a mean speedup of 2.1x (max 2.7x) on a GPU, and 1.3x mean speedup (max 1.9x) on the CPU, while staying within 1 percentage point of inference accuracy loss. For two different accuracy-prediction models, ApproxTuner speeds up tuning by 12.8x and 20.4x compared to conventional empirical tuning while achieving comparable benefits. In the talk, I will highlight applications from the domains of connected self-driving vehicles and agricultural robotics that can leverage our optimizations.

ApproxTuner is developed on top of HPVM, a new general compiler infrastructure for heterogeneous systems developed at UIUC. This is a joint work with some of the original designers of the LLVM compiler. ApproxTuner has been presented recently at PPoPP 2021 and my previous works are available at
https://misailo.cs.illinois.edu/

Регистрациона форма за учешће на Семинару је доступна на:
https://miteam.mi.sanu.ac.rs/asset/CW5nJWDSEZDj7p32p

Пренос семинара је доступан и регистрованим и нерегистрованим корисницима на следећем линку:
https://miteam.mi.sanu.ac.rs/asset/4LNW8WtML7rLKojoz