2017 ITRI Tech day 阙志克

Title; from next generation smartphone to AI

share of smartphone market profit, iphone takes 92, samsung 15, bb 0, ms neg 4, lenovo neg 1percentage

android phones are becoming dumb devices.

apps are revenue n profit resource.

there are 3 topics to discuss today.

1.how can smartphone vendors know more about their users  by big data like google, facebook?

2.How to embedded apps without install?

3.how can APlize app? 

1. know more about users

know more about the IoI, latest items of interests,

how to capture the interests of SP users by data?

How to capture the textual inputs and context of SP?

How to influence users? recommend?

2. App fatigue

85 percentage time on top 3 apps

50 top1, 18 2, 10 3, 5 for the other 2

no app for future sp

install one app to replace the current app, 

it is called app streaming, to run all existing android apps.

instantly run an app without having to download.

its benefits, security n privacy, app trial not to install, market place use not to install for a short time or once.

Solution,

hardware. ARM SoC

software. vitualization options,

android vm, container, serverless computing model

3. Aplize sp smartphone apps

extract info from apps,

add features to apps

aggregate multiple apps to perform on an app

example,

bank apps are independently developed and no interactive.

new app to integrate all bank apps.

bot interface to integrate different messengers, and back end application.

第二个主題 AI. deep neutral network dnn based machine learning.

Why AI will be successful? Because

1.algorithm breakthru, enables training of dnn

2.large high quality training data set,  eg imagenet

3.availability of high performance GPU

what Taiwan can do on AI?

Apply AI to improve the value of different existed industries

How to develop AI to build a new industry by AI

Direction..

1.DNN model training, 1.model quality like tranfer learning, new training tech(googlenet, reinforcement learning etc、). 2. training speed, minimum training round numbers (減少总回合数) and reduce the overhead of each training round(減少每回合的時間消耗)

2. Dnn training data collection, 

imagenet is a successful case

training data sourcing

crowd sourcing

human based computation, image labeling

semi automatic, label propagation

automatic generation, generative adversarial network gan or training scenario synthesis 現有情境

unsupervised learning

it is easy to collect data such as retail spending like consumer invoice, street views, auto driving, restaurant info

3. dnn interface engine

cloud based dnn inference processor. like google tpu 

embedded dnn i. p. nvidia ot intel

dnn complier, from dnn model to executable code.

next major battleground, vehicle computing

4. human competitive AI system

ITRI focus on

autonomous driving

natural language learning

from analysis to synthesis

summary

dnn is expected to be the focus

dnn i.p.

dnn i. subsystem

dnn training model

dnn training data collection