AI-as-a-Service on Blockchain

�Steve Liao

​

12小時前寫完的code大家已經在用了：�

FinTech 世界：All-time, All-place, All-provider: �高度競爭，用code 改變世界者的世界。

不再是“有想法,沒辦法”的世界

​

文化最重要：�做中學 玩真的: Tech�跨領域: Fin

廖世偉博士

• 廖世偉博士的志業是通過創新，產品，Open Source 來教育服務, 改變世界。
• 廖博士 (PhD Stanford) 在斯坦福大學、英特爾、谷歌公司工作了22年，獲得谷歌內部頒發的最高榮譽獎：創始人獎（Founder's 獎）。
• 2013年廖博士從谷歌退休回到斯坦福大學教授程序分析和優化，也在台灣大學教授 最新的安卓系統和金融科技大數據課程。
• 天龍八部（such as 臺大幫幫忙) on Gcoin 區塊鏈。
• 台大黑客松：在天龍八部之前的創新機制，與時俱進。

General AI vs. Specific AI

• 不只做 specific domain 的 machine learning，而是學到 general 的知識
• 就像 Google 一樣：AlphaGo and DeepMind examples
• General AI 仍然需要透過某個 domain 來驗證
• General AI 通常會找很難的領域，例如 playing GO
• General AI: 研究方向應該朝著讓 Machine自己解出方法來而不是靠人類告訴他該怎麼擷取。
• Specific AI vs. General AI:�Feature selection vs. Genuine Machine intelligence
• Domain expert vs. Computer Science

​

​

AlphaGo

• Policy Network
• Value Network
• Monte Carlo
• Re-enforcement learning

AlphaGo experience is General AI

• It sheds light on what is intuition, what is intelligence, what is insight.
• 神來之筆

AlphaGo of 2016 vs. Deep Blue of 1997

• General Intelligence vs. Brute Force
• AI vs. Hardware game
• Intelligence vs. Alpha-beta search

AlphaGo experience is General AI

• AlphaGo sheds light on what is intuition, what is intelligence, what is insight.
• 神來之筆

母雞帶小雞

• General AI vs. Specific AI

Agenda

• Specific AI case study:
• Our deep learning for MRI-analytics-and-prediction
• Domain knowledge and feature selection is key
• Specific AI + General AI case study:
• Our 理財機器人

Our 理財機器人

• 我們理財的算法主要分兩部分，OLMAR跟RNN。
• OLMAR algorithm requires domain knowledge to perform portfolio selection.
• OLMAR 將財務知識演算法化
• Our RNN
• Leveraging Google’s Tensorflow
• http://colah.github.io/posts/2015-08-Understanding-LSTMs/
• Deep Learning 的部分是在RNN，RNN還是 general的解法，沒有用到 domain knowledge。

Our 理財機器人

• 研究方向主要是希望能利用 domain knowledge 來將 data做初步處理 (OLMAR部分)
• 再將處理過後的 data交給 RNN : Deep learning部分
• Generalization: 只要其他 scenerio的 data可以透過演算法整理成 sequence的形式，就能將 data用 deep learning處理

理財機器人’s trends: Chatbot and 深度理財

• Chatbot: Virtual assistant
• Key technologies:
• Natural Language Processing
• Self-Learning
• 深度學習 for 理財

Chatbot

12

Every business should have an APP

Every business should have a APP

Chatbot

#1 Access the 1 billion users platform

Over 1 billion users

Over 220 million MAU

67 % said they expect to message with business in next 2 years

53 % said they are more likely to buy with a business they can message directly

#2 Everyone hates the bothersome navigations

83% think it’s hard to find what they want efficiently in the bank website

Hi, I’d like to apply for a credit card

Dear Kevin, here is the application form for you.

Please fill in and return to us.

#3 Everyone needs a personalized service

Pros:

​

• Connected with the bank service
• Access account information in chat interface

​

Cons:

​

• Lack of AI
• Only support naïve commands (1, 2,…)

​

Chat Platforms

​

Facebook, Line, SMS…

Natural Language Understanding

Cooperate with

Bank API

Natural Language Generation

Our AI Technology

“How much can I spend on my credit card?”

Intent: “ASK QUOTA”

Quota: $200

Hi David, you have $200 to spend before 1st Dec. Happy to organize a credit limit.

Demo

19

深度學習 for 理財機器人

20

深度理財 Outline

• Overview
• Problem Formulation
• Related Work
• Methodology
• OLMAR
• RNN
• OLMAR+RNN (LSTM)

21

深度理財 Outline (cont.)

• Experiments
• Data Description
• Platform
• Result
• More Applications
• Future Work
• Summary of 深度理財

22

Price Prediction

• Prediction on price changing for improving financial investment has been a hot topic for recent years.
• There are a bundle of methods claimed to be efficient while being used on predicting price of stocks or other financial merchandise.
• Several algorithms are then introduced to help predict price trend.

23

OLMAR Introduction

• We want to know that these algorithms are really useful or not. Even if they are, is there any possibility to make it even better?
• OLMAR algorithm is a widely-recommended algorithms and it does have a good performance on helping people pick the right portfolio.

24

100% Return Rate Possible?

• Several algorithms also seem to be efficient.
• Our work is to try to compare OLMAR with other algorithms, and furthermore, combine it with Deep Learning.
• If it works, it can help us get 100% return per year (while neglecting transaction fee)

​

25

Portfolio Selection & Online Learning

• Portfolio selection stands for a combination of several financial investment.
• As the saying goes, “don’t put all your eggs in one basket”.
• Research on portfolio’s behaviour is therefore more attractive than that on a single stock.
• Also, since the data come in a sequential order, online learning fits well in the scenario.

26

Problem Formulation (1/3)

• Here’s our parameter setting: On the t th period, the assets’ prices are represented by a close price vector p(t) ∈ R
• Each element p(t,i) represents the close price of asset i.
• The price changes are represented by a price relative vector x(t) ∈ R, and x(t,i) = p(t,i)/p(t−1,i).

27

Problem Formulation (2/3)

• An investment on the t th period is specified by a portfolio vector b(t) = (b(t,1), ..., b(t,m)), where b(t,i) represents the proportion of wealth invested in asset i.
• We set b(1) = (1/m)*1, where m is the amount of assets.
• On the t th period, a portfolio b(t) produces a portfolio period return s(t), that is, the wealth increases by a factor of s(t) = b(t)^⊤ * x(t) = ∑(m, i=1)b(t,i)*x(t,i).

28

Problem Formulation (3/3)

• GOAL: Optimize the final cumulative wealth, which is the S(T).

29

Related Work (1/2)

• Empirical results show that mean reversion, which assumes the poor stock may perform well in the subsequent periods, may better fit the markets.
• Indeed, algorithms based on this assumption has good performance on specific datasets.

30

Related Work (2/2)

• However, those algorithms[2][3] do not perform well on most datasets, since they only rely on a simple assumption that the predicted next price relative x˜(t+1) will be inverse proportional to last price relative x(t).
• We got two key points: poor stock has a higher possibility to perform well in the subsequent periods, but at the same time we can’t simply assume that x(t+1) will be inverse proportional to x(t).

[2]: Bin Li. Passive Aggressive Mean Reversion

[3]: Bin Li. Confidence Weighted Mean Reversion

31

Methodology-OLMAR (1/8)

• On-Line Moving Average Reversion (OLMAR)[4] algorithm
• Based on the conclusion we got on the previous page, On-Line Moving Average Reversion(OLMAR) algorithm presents several new concepts.
• The essential idea is to exploit multi-period moving average (mean) reversion.

​

[4]: Bin Li. On-Line Portfolio Selection with Moving Average Reversion. 2012

32

Methodology-OLMAR (2/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• Rather than p˜(t+1) = p(t−1), OLMAR assumes that next price will revert to Moving Average (MA) at the end of t th period
• That is, p˜(t+1) = MAt (w) = 1/w*∑(i=t, i=t−w+1) p(i) , where MAt(w) denotes the Moving Average (MA) with a w-window

​

33

Methodology-OLMAR (3/8)

​

• On-Line Moving Average Reversion (OLMAR) algorithm
• We propose a price relative vector following the idea of “Moving Average Reversion” (MAR), which is as shown below:

x˜(t+1, w) = MAt (w)/p(t) = 1/w*(p(t)/p(t) + p(t−1)/p(t) +···+ p(t−w+1)/p(t)) = 1/w(1 + 1/x(t) +···+ 1/⊗(w−2, i=0)*x(t−i))

34

Methodology-OLMAR (4/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• Based on the expected price relative vector, we further adopt the idea of an effective online learning algorithm PA[5].
• PA stands for Passive Aggressive learning.
• If the classification is correct, PA passively keeps the previous solution
• If the classification is incorrect, it aggressively approaches a new solution

​

[5]: Crammer, K. Online passive-aggressive Algorithms

35

Methodology-OLMAR (5/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• Part 1.

​

36

Methodology-OLMAR (6/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• Part 2.

​

​

​

​

​

[6]: J. Duchi. Efficient projections onto the l1-ball for learning in high dimensions. In ICML’2008

​

37

Methodology-OLMAR (7/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• The solution of OLMAR without considering the non-negativity constraint is b(t+1) = b(t) + λ(t+1)*(x˜(t+1) − x¯(t+1)*1)
• Note that x¯(t+1) = 1/m*(1* x˜(t+1)) denotes the average predicted price relative and λ(t+1) is the Lagrangian multiplier calculated as below

λ(t+1) = max{0, ϵ − (b(t)*x˜(t+1))/(∥x˜(t+1)−x¯(t+1)*1∥^2)}.

38

Methodology-OLMAR (8/8)

• On-Line Moving Average Reversion (OLMAR) algorithm
• Empirical observations of the parameter sensitivity show that the final performance is sensitive to the parameter w.
• For each period, we treat the individual OLMAR with a specified w ≥ 3 as an expert and combine multiple experts’ portfolios weighted by their historical performance.

39

Methodology-RNN (1/3)

• What makes Recurrent Networks so special?
• From sequence to sequence

​

​

​

​

�from: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

40

Methodology-RNN (2/3)

• How to deal with long-term memory
• LSTM

​

​

​

​

​

from: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

41

Methodology-RNN (3/3)

• Trial #1:
• using x(t-n)~x(t) to predict x(t-n+1)~x(t+1)
• REAL VALUED
• Trial #2:
• Classify each x(t) into two categories:
• if x(t)>1 then x(t) = 1
• else x(t) = -1
• using x(t-n)~x(t) to predict x(t-n+1)~x(t+1)
• CATEGORICAL

​

​

​

​

​

42

Methodology-Combined (1/2)

• Trial #1:
• First half: OLMAR+training LSTM
• Second half: predicting LSTM

43

Methodology-Combined (2/2)

• Trial #2:
• First half: OLMAR+training LSTM
• Second half: Using predicted labels as the sign to whether or not follow the mean reversion.
• if x(t+1) = -1:
• if MA(w)/p(t) > 1: x(t+1) = p(t)/MA(w)
• else: x(t+1) = MA(w)/p(t)
• if x(t+1) = 1:
• if MA(w)/p(t) > 1: x(t+1) = MA(w)/p(t)
• else: x(t+1) = p(t)/MA(w)

44

Insight

• 如果 Specific-Domain-AI 說要跌，但 General AI (RNN) predicts 要漲，這時是一種 intelligence 的可能。
• Follow General AI 的方向：改成要漲。這是 Passive Aggressive 中的 Aggressive 部份。
• 漲幅還是靠 OLMAR+，只是換個方向。
• 如果 Specific AI and General AI agree with each other，then 就用 Specific AI 即可。
• 這是 Passive Aggressive 中的 Passive 部份。�
• In a way, our OLMAR+RNN is similar to that part of AlphaGo design.

45

Experiments

46

Experiment (Data Description)

47

Experiment (Platform)

• Platform
• We use Python code to construct the whole project.
• Library we use:
• Tensorflow (for training RNN)
• Numpy, Pandas (for developing the trading system)
• Machine: GTX980 with 4GB RAM

48

Experiment (Result trail #1 1/3)

Dataset: NYSE(O)

FAIL! Because the given features are not strong enough to give the exact price prediction

49

Experiment (Result Trial #2 2/3)

Dataset: NYSE(O)

50

Experiment (Result Trial #2 3/3)

Dataset: NYSE(N)

51

Experiment (appendix)

Parameters for LSTM with 14 epoches

Around 2 Hrs on training

52

More Applications

• For stock market trading, the combined algorithm is applicable for high frequency trading. e.g. training at night and prediction in the trading period.

53

Future Work

• Combine transaction fee
• More dataset to prevent survivorship bias
• Parameter tuning for LSTM

54

Summary of 深度理財

• Our proposed algorithm pays attention not only to the mean reversion part (follow the loser), but also mean advancement (follow the winner).

​

• So far, the performance is better than OLMAR but the computational time is a lot longer.

55

AI-as-a-Service Edge Platform (Cognitive Edge Platform) on Blockchain Platform

57

Blockchain Platform for Cognitive Edge Computing

Edge Computing for IoT devices

Edge Computing for IoT devices

58

Cloud Computing

Edge Devices in Edge Computing

59

• Computing offload
• Data caching/storage
• Data processing
• Distribution request
• Service delivery
• IoT management
• Privacy protection

Collaborative Edge

60

• Edge devices can collaborate in local　

by predefined API protocol between edge devices

​

• Information can be fused for synergy and efficiency

​

• Example:
• Government
• City/country-level issues
• Disaster management

​

• Company (HQ, branch)
• Management in different sites

​

​

W. Shi, J. Cao, Q. Zhang, Y. Li and L. Xu, "Edge Computing: Vision and Challenges," in IEEE Internet of Things Journal, vol. 3, no. 5, pp. 637-646, Oct. 2016.

61

Blockchain Platform for Cognitive Edge Computing

Cognitive Edge Computing

Deep Learning (Example: Google FaceNet)

62

Deep

Learning

​

​

​

inference

models

models

967A48 = Alan

01A611 = Adam

983400 = Betty

96B259 = Barbara

Face Feature Vector List

Deep Learning in Cognitive Edge Computing

63

IoT

device

• Privacy Concern
• Data Transfer Issue
• Network Issue

Example of Model Size (Google FaceNet)

64

Incremental Learning

on Edge Device

REF:　Florian Schroff, Dmitry Kalenichenko, James Philbin, “FaceNet: A Unified Embedding for Face Recognition and Clustering” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 2015. Cite as: arXiv:1503.03832 [cs.CV] (or arXiv:1503.03832v3 [cs.CV] for this version), Submitted on 12 Mar 2015 (v1), last revised 17 Jun 2015 (this version, v3), https://arxiv.org/pdf/1503.03832v3.pdf

Collaborative Edge in Cognitive Edge Computing

65

Collaborative

Edge

deep

learning

models

HQ

Oversea Branch

Hotel

Example :

Collaborative Edge for a Company

66

Company HQ

CEO

RD

OP

HR

Manager

MKT

Visitor

Visitor

Oversea Branch

Hotel

Company Guest

Regular Guest

Requirement for cognitive edge computing

67

68

Blockchain Platform for Cognitive Edge Computing

Blockchain as de-centralized database

Brief introduction to Blockchain

• Blockchain is from Satoshi’s paper in 2008:

“Bitcoin: A Peer-to-Peer Electronic Cash System”

• Blockchain is a trust machine that changes how the world functions

(Economist 2015-10-31)

​

• Blockchain can become the ledger in the internet era

(Wall Street Journal 2015-1-25)

• By 2027 10% of GDP is on blockchain

(World Economic Forum, Davos 2016)

69

Blockchain

70

Immutable

Security

Transaction

is Settlement

Traceable

Ver.

How a �Blochchain�Works

71

Source: Financial Times

• “Wallet” is an account to store records

​

• “Coin/Token”

are resource can

be transferred

​

How blockchain keeps records Immutable

72

Block Header

Hash

Previous Block Header Hash: A hash in internal byte order of the previous block’s header. This ensures no previous block can be changed without also changing this block’s header.

Merkle Root Hash: The merkle root is derived from the hashes of all transactions included in this block, ensuring that none of those transactions can be modified without modifying the header.

Target nBit: An encoded version of the target threshold. This block’s header hash must be less than or equal to.

Nonce: An arbitrary number miners change to modify the header hash in order to produce a hash below the target threshold‧

txn_count: The total number of transactions in this block

Txns:　Every transaction in this block, one after another, in raw transaction format.

Block Header Hash

Hash of Block Header

Block Header Hash

Hash of Block Header

Ver.

Ver.

What kind of Trust Machine is it?

73

Trust Machine

supports

• Privacy, Security: CoinJoin, CoinShuffle, Zero-knowledge proof, Coin Mix, Mix server, 51% attack
• Math: One way function
• Crypto: Hashing/Signature/ECC/SHA-256
• Economics: Monetary policy/Game theory
• Consensus: Byzantine Generals problem/Hashcash
• Scalability (Global)
• Flexibility (Governance structure, smart contract)

State machine-based immutable decentralized Trust Machine

74

state

state

state

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo

utxo: unspent transaction output

How the Internet and blockchain both work. Ledger is synchronized by “full” (red) nodes.

75

Blockchain Platform for Cognitive Edge Computing

Gcoin Blockchain

Consortium/Permissioned Blockchain

76

Source: Financial Times

Gcoin Blockchain

77

78

79

80

81

82

alliance

member

alliance

member

alliance

member

issuer

issuer

issuer

issuer

83

Blockchain Platform for Cognitive Edge Computing

Gcoin Blockchain Platform

for Edge Computing

Gcoin Blockchain Platform for Edge Computing

84

Edge Computing

Management

Gcoin

Blockchain

Management

Token: resource records

Smart contact : resource handling

85

Hospital2

Government

Alliance Member

Medical

Alliance Member

Rescue system

Alliance Member

Approval process

Token Issuer

Financial

Token Issuer

Rescure resource

Token Issuer

Medical resource

Token Issuer

Ministry of

Health and

Welfare

Hospital1

Hospital3

City2 Gov

City2

Finanal Dept.

Collaborative Edge Devices

Gcoin Blockchain Platform for a Government

86

Collaborative Edge Devices

Company Management

Alliance Member

Human Resource

Alliance Member

Supply Chain Alliance Member

Approval process

Token Issuer

Financial Token Issuer

Supply Chain Management

Token Issuer

Employee Management

Token Issuer

Face

Token

Gcoin Blockchain Platform for a Company

“Face Token”

87

Company HQ “wallet”

CEO

RD

OP

HR

Manager

MKT

Visitor

Visitor

Oversea Branch “wallet”

Hotel “wallet”

Company Guest

Regular Guest

“Smart contact”

of Gcoin blockchain

to handle management

Use Token management for collaborative Edge devices management

Gcoin Blockchain as Platform for Cognitive Edge Computing

88

Summary of Blockchained AI

• Cognitive Edge Computing
• For privacy concern, incremental models can be learned on edge devices.
• A solution is needed to management “incremental models” among a group of edge devices
• De-centralized, Secure, Traceable, Immutable, Efficient, Flexible for configuration

​

• Blockchain Platform for Collaborative Edge
• Blockchain is a candidate solution
• Blockchain can be a de-centralized, secure, traceable and immutable solution
• Gcoin blockchain
• Gcoin blockchain provides more advantages of efficiency, flexibility for configuration and management

​

• Use Gcoin Blockchain for cognitive edge computing for collaborative edge
• Transfer edge device management problems to Gcoin “Token” management
• For example, a new “Face token” for incremental models of face recognition
• Smart contract can be used for automatic execution

​

89

數位經濟美麗新世界的展示

區塊鏈點數平台

• 各機構間使用區塊鏈，作到高效、快速的交易共同信任機制
• 第三方支付帳聯網
• 一般用戶可在行動/第三方支付系統間互聯互通
• 銀行/業者可快速連結帳聯網，享有互聯互通帶來的網路效應，將資源專注於提供創新用戶體驗與服務
• 供應鏈金融帳聯網
• 營運資金和現金流量的實時監控和管理
• 實現營運資金及現金流的靈活調度

帳聯網 API

數字資產交易平台

• 具有強大可追蹤性與延展性的數字資產商品交易平台
• 實物商品：電影票、演唱會、藝術品
• 非實物商品：貸款商品、股權商品、債權商品、智慧財產權
• 平台運營商 (Platform operator)、網路商家 (Merchant) 互聯互通
• 提供平台運營商互相加盟的環境，例如：票據投融平台。提供網路商家可同時在多個平台運營商之間獲得客戶

​

​

區塊鏈：點數管理平台及交易平台

• 區塊鏈數位經濟還有大數據好處
• 大數據預測分析 – 依據用戶行為做定向廣告
• 商務情報 – 提供對業務經理人具洞見的指標指數

GCOIN 區塊鏈使用情境

私募平台

票據平台

去中心化交易所

crowdsourcing平台

顧客忠誠計劃

支付結算系統

95

THANK YOU!

Blockchain Platform for Cognitive Edge Computing

Blockchain: �State-Machine based Trust Machine

Remember: blockchain machine is Mathematics-�based:

​

• Math just works;

You don’t need to trust it:

Trust Machine vs. Trust Company

What kind of Trust Machine is it

• To database researchers:
• Blockchain is NOT just a brand new data store.
• To computer network researchers:
• Blockchain is NOT just a P2P network.
• To virtual machine and compiler professors:
• Blockchain is NOT just a smart contract.
• To programming language professors:
• Blockchain is NOT just a programming platform.

​

Blockchain’s key technology: State-Machine based immutable Trust Machine

• Don’t jump to database, network, … applications.

Data structure of Blockchain

Block Header

Hash

Previous Block Header Hash: A hash in internal byte order of the previous block’s header. This ensures no previous block can be changed without also changing this block’s header.

Merkle Root Hash: The merkle root is derived from the hashes of all transactions included in this block, ensuring that none of those transactions can be modified without modifying the header.

Target nBit: An encoded version of the target threshold. This block’s header hash must be less than or equal to.

Nonce: An arbitrary number miners change to modify the header hash in order to produce a hash below the target threshold‧

txn_count: The total number of transactions in this block

Txns:　Every transaction in this block, one after another, in raw transaction format.

Block Header Hash

Hash of Block Header

Block Header Hash

Hash of Block Header

“Mining” Blockchain (Proof-of-Work)

Block n

Block Header Hash

Block (n-1)

Block Header Hash

Block (n-2)

Block Header Hash

Block Header Hash

Current Blockchain

New Block (n+1)

Block Header Hash

< Target nBits ?

Set

new

Nonce

value

No

Yes

Successful Mining

Hash of Block Header

Distributed Mined Blockchain to Peer

A

B

C

D

E

F

G

H

Block (n+1)

Successful Mining

Arriving at Consensus�

• Although the accepted chain can be considered a list, the block chain is best represented with a tree.
• The longest path represents the accepted chain.
• A participant choosing to extend an existing path in the block chain indicates a vote towards consensus on that path. The longer the path, the more computation was expended building it.

101

Model Distribution (1) for Cloud Server

102

Peer-to-peer download

Centralized download

models

models

Model Distribution (2) for Edge Device

103

Peer-to-peer download

Centralized download

105

Backup Slides