Compiler-Level Approximate Mobile Computing

Fabjančič, Matevž (author), Pejović, Veljko (mentor)

With the widespread use of smartphones, wearable devices and many applications of deep learning (DL), there is a growing interest in deploying DL on low-power devices. However, due to inferior computational resources and battery capacity limitations, this is a challenging task. One solution to this problem stems from approximate computing – by using approximations, we can sacrifice accuracy for better energy-efficiency. We develop an end-to-end system for adaptive approximate mobile computing (AMC), which enables transforming high-level definitions of convolutional neural networks into approximable DL models suitable for deployment within Android applications. We define ways of adaptively selecting among approximation levels to achieve better energy-efficiency of DL on smartphones while preserving the option of using non-approximated neural network variants. We evaluate the benefits of adaptive AMC on a concrete use case.

• Language: English
• Keywords: Ubiquitous Computing, Adaptive Approximations, Approximate Computing, Compilers, Android, Deep Learning
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2021
• PID: 20.500.12556/RUL-132867
• COBISS.SI-ID: 84027395
• Publication date in RUL: 05.11.2021

Secondary language

• Language: Slovenian
• Title: Približno računanje na ravni prevajalnika na mobilnih napravah

Abstract

Razširjenost pametnih naprav, kot so telefoni in ure, skupaj z množično uporabo aplikacij globokega učenja kličeta po uporabi modelov globokega učenja na napravah nizke zmogljivosti. Zaradi računske zahtevnosti globokega učenja pa je to na napravah z omejenimi računskimi viri težko izvedljivo. Z vpeljavo približnega računanja v modele globokega učenja lahko za ceno natančnosti modelov prihranimo na porabljeni energiji. V tej nalogi razvijemo enoviti sistem za prilagodljivo približno računanje na mobilnih napravah. Sistem omogoča, da visoko nivojske opise nevronskih mrež pretvorimo v modele z nastavljivo približnostjo in jih uporabimo v aplikacijah za naprave z operacijskim sistemom Android. Na primeru uporabe pokažemo, da lahko z različnimi sistemi samodejnega prilagajanja približnosti modelov dosežemo boljšo energijsko učinkovitost modelov na mobilnih napravah in ohranimo možnost klasifikacije z neaproksimirano nevronsko mrežo.

• Keywords: vseprisotno računanje, prilagodljivo aproksimiranje, približno računanje, prevajalniki, Android, globoko učenje