applications on ai · alexnet, 2012 first breakthrough deep learning which excels statistical...

39
Applications on AI DATE SUMMIT 2018 Hajime Hotta

Upload: others

Post on 25-Mar-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Applications on AI DATE SUMMIT 2018

Hajime Hotta

Page 2: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Myself

�2

Hajime Hotta, Ph.D — Computer Science Ph.D at the age of 25 Having 2 tech-driven companies successfully exited [email protected]

Early Stage AI InvestorAI Startup

HQ in Tokyo, Japan

Page 3: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

What is AI?AI: Capability of automating intuitive process

�3

ICT

Information and Communication

Technology

AI

Artificial Intelligence

Fully-defined Rules Intuitive Processes

Majority of the world’s issue are NOT easily defined by the strong

rules which ICT requires

AI technologies unlock the automation of many processes

which are intuitive yet not-definable by rules

Page 4: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Intuitive Process: Cat & Dog classificationAI: Capability

�4

Page 5: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Difference among Rule-based, ML and DLDeep Learning has capability of NOT even considering the features for algorithms to pay attention to

�5

Rule-based Approach Traditional Machine Learning (ML) Deep Learning (DL)

Classify

Define the rule 1) Cat’s ears are more XXX 2) The shape of mouth is XXX

TOO MANY EXCEPTIONS

ML Specialists extract the points of attention (features)1) Mouth shape 2) Ear shape 3) Color => Machine Learning to define the threshold to judge if it’s cats or dogs

Need experts

Almost 100% Automated Algorithm itself automatically extract features, then automatically Finetune the model for better accuracy

Page 6: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Three Technical Functions of AI(1) Recognise (2) Predict (3) Prescribe

�6

Recognise Predict Prescribe

Autonomous Car Industrial Automation

Image / Speech Recognition Natural Language

Understanding, etc.

Financial Prediction Machine breakdown

Prediction …

Prediction based on the recognition

of past patterns

Prescription based on the

predicted situations

What Deep Learning Can essentially do

Page 7: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Deep Learning being improvedThe deep learning technology gets very fast improvement

�7

0.05

0.10

0.15

0.20

0.25

0.30

2010 2011 2012 2013 2014 2015 2016 2017

AlexNet

ZFNet

GoogLeNet

ResNet

Ensemble

SENet0.036 0.030 0.023

0.070

Classification Error of ILSVRC

Breakthrough 1

Breakthrough 2

Page 8: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

AlexNet, 2012First breakthrough deep learning which excels statistical methods

�8

- 8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980) - Avoid overfitting problems by DropOut (randomly invalidate neutrons inside) - ReLU function

First high-performance Deep Learning

Page 9: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

ResNet, 2015Second breakthrough deep learning; most of the latest models are the modification of ResNet

�9

- Shortcut mechanism to allow the training process to go deep => Enable very deep network

- Batch Normalisation to stabilise and accelerate training processes

34 layersMajorities of DL models are

based on ResNet architecture

Page 10: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

SqueezeNet = Lightweight Deep LearningSqueezeNet, a very lightweight model with AlexNet level accuracy

�10

Very small model size with 1x1 Convolutional Layer (Fire Module) => 50x smaller parameters, <0.5M network size => Workable with Raspberry Pi level Network => Accuracy is AlexNet level

Breakthrough for low-spec hardwares

Page 11: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

LSTM = Good for time-sequence dataLong Short term Memory; A better version of Recurrent Neural Network

�11

- Proposed originally in 1997, but with the combination of convolutional neural network, LSTM works very well in many applications

Neural Network model applicable to Time-Sequence data (Success in Speech Recognition, Natural Language Processing, Machine Translation, etc)

Base Technology for Speech Recognition and Natural Language Processing

Page 12: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Attention ModelAn upgrade technology of RNN

�12

A mechanism to Pay Attention to what really matters

Successful in dealing with long, complicated sentences

Empower LSTM to analyse more complicated data

Page 13: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

(2017) Generative Adversarial NetworkAlso known as GAN; a technology to generate new realistic data

�13

Two networks combined for the content generation

Key technology to

GENERATE data

(Data Synthesis)

Page 14: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Data SynthesisData Synthesis using Deep Learning models, GAN, etc.

�14

Human Speech

Human Handwriting

Musics Jazz

Natural Language Generation

Human Illustration

Arts

Human Conversation

Drone Captured Photos 3D

Spaces

Objective 1

Data generation for customer-facing functions - Computer-generated news and curations - Computer-generated voice for robot speaking

Objective 2

Data Generation to give more training data for deep learning to achieve further accuracy - Interpolated photo data for better person tracking

Page 15: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

(2017) Capsule NetworkCapsule Network, proposed by Professor Hinton

�15

Capsule Network Orignial MLP

- Vector input - Affin Transform before neutron - Use of Squash function instead of original

ReLU unit - Removal of bias (+1) - Vector output

Mechanism to preserve the spacial relationship between components

Robust to rotations and scales

LESS DATA

Page 16: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Less Data Trends“Less Training Data” era is beginning => Unlocking many AI with non-big-data

�16

Big trends on Data Synthetics (by GAN, etc) Some Robust Machine Learning Techniques (Capsule Networks)

“Less Data” training is beginning

Many niche AIs will be unlocked

Page 17: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Startups’ technical strategyGo domain-specific, relying on small infrastructure, focusing on less data

�17

Suggested approachGoogle approach

General Domain Specific

Centralised cloud Distributed (on-premises)

Big infra Lightweight / PC-level

All-in-one AI engine Module plug-and-play

Big Data Less Data*

Page 18: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

AI business categorisationFour categories (1) Internet (2) Business (3) Recognition and (4) Autonomous AIs

�18

AI Companies

Internet AI

Business AI

Recognition AI

Autonomous AI

CyberSecurity

Medical Research

Apply AIs to serve users the web services / mobile services with better UX or better conversions

Apply AI to help businesses make decisions, understand customers, marketing, or reduce costs and workloads with AI automation

AI for voice / image / video recognition to extract information and convert into usable data for other systems.

AI to automatically move; Autonomous car, Robotics, Manufacturing automation

CyberSecurity innovations are accelerated by AI as well

Medical / Pharmacy Researches are unlocked by the application of machine learning techniques.

Top Four categories are defined by Kai-Fu Lee ( AI in China: Cutting Through the Hype)

Developer Platform

Enable developers to build ML applications easily / maintain applications easily

Page 19: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Well-Funded Startups in the four areasDiversity of industries are now adapting to AI

�19

Internet AI

Business AI

Recognition AI

Autonomous AI

ByteDance News App $1800M

SenseTime Facial recognition

and Self Driving Car $637M

Face++ Facial

Recognition $608M

DataMiner Total Data

Mining Solution $577M

Zoox Autonomous

Car $290M

InsideSales.com

Sales Acceleration

$290M

Equipment controls $258M

Mobvoi Smartwatch

In China $257M

Financial Advisory App $204.5M

Intelligent Enterprise Knowledge Search

$170.7M

C3 IoT PaaS for IoT Apps

$103.4M

Page 20: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Other AreasCyber Security / Medical / Pharmacy

�20

CyberSecurity

Medical / Pharmacy Research

Developer Platform And AI DevOps

$297M $182.3M $106M $481M

$313M $202M$196M $85M

$73.6M $52.93M $46.9M $22.7M

Page 21: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Internet AIApply AIs to serve users the web services / mobile services with better UX or better conversions

�21

Area of Startup Opportunities Personal Loan: Big-data-driven Credit Scoring to micro -lending Telehealth: Allow patients to communicate with doctors with chatbot experiences Adaptive Learning: For students / learners to learn effectively through AI-based personalisation technologies Robot Advisor: Financial Planning advisor done by AI. Applied to insurance plan advisors, too News App: News curator application personalised by AI

Page 22: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Business AIApply AI to help businesses make decisions, understand customers, marketing, or reduce costs and workloads with AI automation

�22

Customer Support

Business Process

Professional Works

Sales Process

Big Data Analysis

Five Areas that attract investors

Document, Knowledge Discovery,

etc.

Call Centre/ Emails /

Chat-based Customer support

Legal Accounting Compliance

Patent Attorney Medical Doctor

Identify the focus customer, Call Logs,

CRM

Abuse Detection Fraud Detection

Customer Identification

Page 23: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Recognition AIAI for voice / image / video recognition to extract information and convert into usable data for other systems.

�23

Smart Speaker Tracking people/Car Industrial Automation

Sensor networks to identify robot breakdown

Camera for Product quality inspection

People tracking in retail Shops

Car tracking for Public sector Biometrics

Smart Speaker In-car Smart speaker

Page 24: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Autonomous AIAutonomous Car + Robotics

�24

Autonomous Car Industrial Automation

Lots of partial Technologies have invented and merged - Tire sensor, - Driver Emotion Detection - Road Quality

Automation of blue-collar works

By robots

Inventory Management

A robot to manage Inventories

Page 25: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

AI Product-Market Fit Four types of B2B demands, several proven areas for B2C business models

�25

B2B

Blue-Collar jobs Replaced with robots

White-Collar jobs Replaced with AI softwares

Minimise the loss of Business opportunities Increase Revenues

Page 26: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

KindRed Solutions — $43MRobotics Arms

�26

Robotics

Page 27: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

C3 IoT — $103.4MPlatform for IoT devices to make the AI easily configurable and controllable through SaaS

�27

Edge Computing

Page 28: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

WorkFusion — $121MRPA (Robot Process Automation) with AI

�28

RPA

Page 29: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Sift Science — $106.6MFraud Prevention (Payment, Content Abuse, Account Takeover, etc. )

�29

Fraud Detection

Page 30: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

InsideSales.com — $264MAn AI-powered predictive sales acceleration platform

�30

Sales Boost Up

Page 31: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

AI-driven Product patterns AARRR framework to identify the channel of users

�31

3) Chatbot for much better UX for conversion

4) Credit Scoring x FinTech

5) Purchase Prediction x Subscription Commerce

2) Customer Personalisation

6) Robots alternative to professionals

7) New Devices for home and wearables

1) Autonomous Car

Page 32: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

1) Autonomous CarAutonomous car is one of the

�32

Mobileye: Acquired by Intel in $15.3B Self-driving car technology kickstarted with

partnership with Tesla (ended) Now embedded in many brands

Nauto: $186M Funded, led by Softbank Automotive technology using sensors and

cameras inside and outside vehicles to prevent accidents

Partnership with car manufactures + Data Collection

Page 33: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

2) PersonalisationPersonalisation no longer means just a recommendation, but a tech to deliver significant benefits to users

�33

Kidaptive: $38.7M Adaptive Learning Technology

(ALT) provider, which employs AI for English Learning

Voyerger Lab, $100M VoyagereCommerce is a

solution for e-commerces to personalise the UX

(recommendation, email, landing page, etc. )

Bytedance, $3110M Personalised news media

empowered by AI

Page 34: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

3) Chatbot for better UXChatbot increases customer engagement significantly

�34

Lemonade — $180M Lemonade is a licensed insurance

carrier that offers homeowners and renters insurance powered by

artificial intelligence and behavioral economics.

Need only 90 seconds to subscribe the insurance

program through Chatbot

Page 35: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

4) FinTechChatbot increases customer engagement significantly

�35

Upstart — $584M Upstart is a lending platform that

leverages artificial intelligence and machine learning to price credit and

automate the borrowing process.

Stash — $116.3M Stash advisor helps guide investors from there, with advice, support and

recommendations.

Page 36: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

5) Subscription CommerceSubscription Commerces are boosted

�36

Stitch Fix, Fashion Stylist, $42.5M Personalised fashion stylist

BirchBox, $10/mo subsciption, $86.9M Personalised mix of cosmetics samples

Page 37: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

6) RobotRobot alternative to professionals such as: Financial Adviser, Medical Doctor, etc.

�37

Robo-Advisor, 204M Wealth Management App

Babylon Health, $85M Chatbot First and connect to real Medical Doctor

through consistent UX

Page 38: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

7) New DevicesRobot Apps

�38

Mobvoi, $256.91M Smart Watch with good designs, China

Deep Sentinel, $7.3M Home Security Device

Page 39: Applications on AI · AlexNet, 2012 First breakthrough deep learning which excels statistical methods 8-8 layered, architecturally inspired by Neocognitron (Fukushima et al, 1980)

Thank you

�39

Hajime Hotta, Ph.D [email protected]