AI Benchmark: Running Deep Neural Networks on Android Smartphones
0. Abstract
In this paper, we present a study of the current state of deep learning in the Android ecosystem and describe available frameworks, programming models and the limitations of running AI on smartphones.
1. Introduction
Android Neural Networks API (NNAPI) [47], designed to run deep learning models on mobile devices.
We present an AI Benchmark designed specifically to test the machine learning performance, available hardware AI accelerators, chipset drivers, and memory limitations of the current Android devices.
2. Hardware Acceleration
2.1 Qualcomm chipsets / SNPE SDK
2.2 HiSilicon chipsets / Huawei HiAI SDK
2.3 MediaTek Chipsets / NeuroPilot SDK
2.5 Google Pixel / Pixel Visual Core
2.6 Arm Cortex CPUs / Mali GPUs / NN SDK
3. Deep Learning Mobile Frameworks
4. AI Benchmark
The AI Benchmark is an Android application designed to check the performance and the memory limitations associated with running AI and deep learning algorithms on mobile platforms.
Test 1: Image Recognition ----MobileNet-V1
Test 2: Image Recoginition ---- Inception V3
Test 3: Face Reconginition ---- Inception-Resnet-V1 network
the first three tests cur-rently represent a core set of architectures for classification problems that are suitable for mobile deployment.
Test 4: Imae Deblurring ---- SRCNN network
Test 5: Image Super-Resolution --- VDSR network
Test 6: Image Super-Resolution --- SRGAN model
Test 7: Image Semantic Segmentation --- ICNet CNN
Test 8: Image Enhancement --- DPED paper
Test 9: Memory Limitations --- SRCNN
Use the TensorFlow Lite library
5 Benchmark Results
5.2 Smartphones and mobile chipsets