Tokyo Tech Homecoming Day "AI and MOT

slide

http://nhiro.org/titech_hcd/page-01.png

Tokyo Tech MOT Homecoming Day "AI and MOT "2018-05-26Senior Researcher, Cybozu LabsPresident, General Incorporated AssociationFUTO DirectorTokyo Tech MOT 2011Admission-2014Completion2018Specified Associate ProfessorNISHIO Hirokazu

http://nhiro.org/titech_hcd/page-02.png

Facebook Group 2Share lecture materials, questions, comments, and anything else.

http://nhiro.org/titech_hcd/page-03.png

Objective 3 of this presentation will increase the probability that people who hear this presentation will create economic value.

http://nhiro.org/titech_hcd/page-04.png

AlphaGo 4First Go Program to Beat the Pros

http://nhiro.org/titech_hcd/page-05.png

Computers Surpass Humans? 5Human Computers

http://nhiro.org/titech_hcd/page-06.png

Deep Learning has shown promise6Vertical Axis: Expectations

http://nhiro.org/titech_hcd/page-07.png

http://nhiro.org/titech_hcd/page-08.png

8 It seems that Deep Learning, a technology like a magic wand, can solve a lot of problems!

http://nhiro.org/titech_hcd/page-09.png

http://nhiro.org/titech_hcd/page-10.png

Parking Difficulty Prediction 10 - Google Maps feature - "How hard is it to park in this area at this time? Easy?" -Deep Learning is amazing!

http://nhiro.org/titech_hcd/page-11.png

We did not use Deep Learning11 "We used logistic regression" Reasons (excerpts) -Easy to understand the behavior -High tolerance to noise -Easy to verify that the model is working properly

http://nhiro.org/titech_hcd/page-12.png

Logistic regression12 -Method created in 1958 -Mathematically equivalent to thin neural nets, not Deep Learning -Will be a dead method, so learning is getting faster

http://nhiro.org/titech_hcd/page-13.png

http://nhiro.org/titech_hcd/page-14.png

Logistic regression 63.5% 14

http://nhiro.org/titech_hcd/page-15.png

The purpose of the means15 -Deep Learning is certainly more accurate than past methods for certain tasks, but it is not a universal tool nor is it a magic wand. Let's improve business management with Deep Learning" is wrong because the tool is decided before the objective is decided.

http://nhiro.org/titech_hcd/page-16.png

AlphaGo 16 - I want you to get a rough sense of the scale - I'll explain in a bit more detail - I won't explain the differences between the detailed versions.

http://nhiro.org/titech_hcd/page-17.png

Amount of data: 28.4 million data from 160,000 games obtained from 17 Go online match servers.

http://nhiro.org/titech_hcd/page-18.png

Learning Results 18Able to guess the human "next move" with a 57.0% correct answer rate.

http://nhiro.org/titech_hcd/page-19.png

Generate data through self-play19Machines play 30 million games against each other because the 160,000 game data generated by human games is not enough (using the characteristics of Go, which allows data to be collected without the intervention of humans or the physical world).

http://nhiro.org/titech_hcd/page-20.png

30 million games is about 1 hour per game for 20 people, so 3500 years (sleepless nights).

http://nhiro.org/titech_hcd/page-21.png

21The exact amount of money spent on the study is, of course, unknown, but based on publicly available information, it is estimated to be about 3 billion yen. 058.html

http://nhiro.org/titech_hcd/page-22.png

Summary 22 - First, collect about 30 million data of "decisions made by humans" - Collect 200 times more data by simulation - Invest 3 billion yen in this process

http://nhiro.org/titech_hcd/page-23.png

23Why did you do it? (Let's think about it from a manager's perspective)

http://nhiro.org/titech_hcd/page-24.png

Why play AlphaGo? 24 - Branding? -To train new employees? -Because it's a world first and it gets employees excited? -To attract investment by showing dreams?

http://nhiro.org/titech_hcd/page-25.png

Investment increases as expectations increase25Vertical Axis: Expectations

http://nhiro.org/titech_hcd/page-26.png

S-curve development of technology 26 bell curve S-curve integral hype curve S-curve development + sustainable growth integral

http://nhiro.org/titech_hcd/page-27.png

Self-fulfillment of the prophecy27 -Promoting the awesomeness of the technology -Expectations for the future are raised -Investment is drawn out -This increases the speed of technological progress.

http://nhiro.org/titech_hcd/page-28.png

28But also counterproductive...

http://nhiro.org/titech_hcd/page-29.png

29 Fear of losing jobs to AI

http://nhiro.org/titech_hcd/page-30.png

30 Knowledge is the Antidote to Fear Ralph Waldo Emerson "Society and Solitude"

http://nhiro.org/titech_hcd/page-31.png

31Based on concrete examples, not vague concepts

http://nhiro.org/titech_hcd/page-32.png

Case Study: Miura City Agricultural Cooperative's Scheduled Dispatch System32 ref: ICT Changes Society Case Study

http://nhiro.org/titech_hcd/page-33.png

http://nhiro.org/titech_hcd/page-34.png

http://nhiro.org/titech_hcd/page-35.png

http://nhiro.org/titech_hcd/page-36.png

Is it bad that "AI is taking away our jobs? Masayoshi Iijima, assistant general manager of the cooperative's sales department, commented positively, "The biggest thing is the reduction in working hours. I have high expectations for the new system because it will reduce the workload of the staff and they can use their free time more efficiently for sales and other activities," he said.

http://nhiro.org/titech_hcd/page-37.png

http://nhiro.org/titech_hcd/page-38.png

Population Change38 Nishio, b. 1973 ('81) I and the generation younger than me grew up in an environment where the number of people decreased every year.

http://nhiro.org/titech_hcd/page-39.png

One in seven will be gone.39 Over the next 15 years, the 20-something population will decline by one in 67.

http://nhiro.org/titech_hcd/page-40.png

If one person in seven disappears...40 The work that is being done by a team of seven now must be done by six people in 15 years if the age structure of the team remains the same If the workload remains the same, the workload per person will increase by 17% for a job that is worked 20 days a month x 8 hours a month with a 17% increase in workload. +27 hours per month. Overtime +1 hour every day or -3 days off every month.

http://nhiro.org/titech_hcd/page-41.png

Maintaining the status quo is 41 "doing the same job the same way we have always done it" is not maintaining the status quo. It is not maintaining the status quo, because the burden on humans will increase and they will not be able to do their jobs. To maintain the status quo, 1/7 of the work must be "taken away by AI" or human capacity must be increased by 7/6.

http://nhiro.org/titech_hcd/page-42.png

42 x AI will take 1/7th of the work ○ AI will take 1/7th of the work off our shoulders ○ AI will be used to augment humans by 16

http://nhiro.org/titech_hcd/page-43.png

Reinvestment of excess time43 -Trend of fewer young people -Hiring new people is slowly becoming more difficult -The idea of "reducing jobs with AI and laying off excess workers" is a bad idea -Reduce jobs with AI and and reinvest the free time created into more productive jobs - Expand and reproduce

http://nhiro.org/titech_hcd/page-44.png

human computer

http://nhiro.org/titech_hcd/page-45.png

human being

http://nhiro.org/titech_hcd/page-46.png

Human + computer = augmented human human human

http://nhiro.org/titech_hcd/page-47.png

Examples of Enhancement47 -Computing power enhancement by computer -Transmission power enhancement by internet -Information discovery power enhancement by search engines Here are some more examples

http://nhiro.org/titech_hcd/page-48.png

Automated Programming48The idea is to automate programming because it is hard to program.

http://nhiro.org/titech_hcd/page-49.png

Before automatic programming49I want to do "add A and B and put them in C." LOAD 1 ADD 2 STORE 3Load the value in memory 1 to the temporary area for calculationAdd the value in memory 2 to itAdd the value in memory 3 to the temporary areaThis is pseudo code to make things simple, In reality, it is much more complicated.

http://nhiro.org/titech_hcd/page-50.png

Automatic Programming 50 "Add A and B to C" is written in the notation C = A + B;. The computer reads this notation and automatically generates LOAD 1; ADD 2; STORE 3; "Formula Conversion System".

http://nhiro.org/titech_hcd/page-51.png

19541957 FORTRAN 51 -FORmula TRANslation system -the beginning of the "programming language" used by almost all modern programmers - a system for converting instructions written in mathematical expressions at a high level of abstraction into instructions written in machine language at a low level of abstraction (automatic programming*). System for converting instructions written in mathematical formulas at a high level of abstraction into instructions written in machine language at a low level of abstraction (automatic programming*) *"The Fortran Automatic Coding System for the IBM 704" (1956)The system was created in 1954, and this manual was created in 1956, and became available to the public in 1956. 56, and it became generally available in 57.

http://nhiro.org/titech_hcd/page-52.png

The Birth of Scripting Languages52 About 30 years ago, languages such as Perl, Python, PHP, and Ruby were born, which are slower than major languages such as C. The execution speed C Python time

http://nhiro.org/titech_hcd/page-53.png

Disruptive Innovation 53 "The Innovation Dilemma," p. 10

http://nhiro.org/titech_hcd/page-54.png

Scripting languages54 -Computers became faster -Slower programming languages could meet customer demand - "slower to execute, but less time-consuming to implement" Scribble-like languages Scribble-like language with "slow execution but little time for implementation

http://nhiro.org/titech_hcd/page-55.png

Impact of Machine Learning55 -The majority of programs are conditional decisions - humans clarify "what the conditions are" and communicate them to the computer - but some tasks are difficult to "clarify the conditions" (Go, image recognition...) -Machine learning is.., ...) - Machine learning is...

http://nhiro.org/titech_hcd/page-56.png

Impact of Machine Learning56Machine learning is a method that allows computers to make conditional judgments by giving them data that represents a situation and "how a human would make a decision in that situation" to learn, without the human having to clarify the conditions. Note: This explanation is limited to supervised learning.

http://nhiro.org/titech_hcd/page-57.png

Reduce cost at the expense of certainty57 - Since humans do not clarify conditions, it is meaningless to ask, "Does it work according to the conditions? The question "Does it work according to the conditions?" is meaningless because the human does not clarify the conditions - "90% of the verification data can be judged as correct" is all that can be said - Reduce the cost of writing the specification at the expense of certainty.

http://nhiro.org/titech_hcd/page-58.png

Summary 58 - Automated Programming - Scripting Languages - Machine Learning Commonality is Reduction of Human Costs

http://nhiro.org/titech_hcd/page-59.png

Academia and Business Perspectives59From a business perspective - the decrease in costs - new goods and their production methods - enter the "commercially feasible range".

http://nhiro.org/titech_hcd/page-60.png

Difference between Academia and Business60Academia - Data is publicly available - Dead technologies have been investigated long ago - Businesses compete for accuracy by inventing new methods - Data is not publicly available Business - data is not publicly available

http://nhiro.org/titech_hcd/page-61.png

Business Evaluation Axis 61Customer ValueNot Newness

http://nhiro.org/titech_hcd/page-62.png

Sub-question 62 for experiencing customer value: A gemstone, which has a 1/2 chance of containing a gem when broken, sells for 20,000 yen. One rough gemstone costs 9,500 yen. Since the gemstone is hard, only 20 pieces can be broken per day.Q1: What is the expected value of revenue that can be earned per day? Purchase cost1/2 gemstone9500 yen x 1/2 sales20,000 yen0 yen

http://nhiro.org/titech_hcd/page-63.png

A1Expected value of earnings63 A1: Since there is a 1/2 chance of getting 20,000 yen, the expected value of earnings per unit is 10,000 yen. 9500 is the purchase price per unit, so earnings per unit is 500 yen. 20 units can be processed per day, so earnings per day is 10,000 yen. Cost 9,500 yen sales 10,000 yen profit of 500 yen per unit.

http://nhiro.org/titech_hcd/page-64.png

Q2Double processing speed device64 Q2: Suppose you have a machine that breaks rough stones at twice the speed of a human. In other words, the number of rough stones that can be broken per day increases from 20 to 40. How much would the expected value of revenue increase?

http://nhiro.org/titech_hcd/page-65.png

A2: The profit of 500 yen per rough stone remains the same. 20 more pieces can be processed per day, so the profit increases by 10,000 yen. Cost 9,500 yen, sales 10,000 yen, profit 500 yen per piece

http://nhiro.org/titech_hcd/page-66.png

Q3 60% discriminator66 Q3: Suppose you can buy a discriminator* that has a 60% chance of hitting a gemstone with a gem in it. How much would the expected value of revenue increase? Purchase cost 60% gemstone 9500 yen x 40% sales 20,000 yen 0 yen* Suppose that by using it before buying a gemstone, you can get a "gemstone with a 60% chance of containing a gemstone" for 9500 yen. We do not care this time about the fact that the shopkeeper may not like it. Also, assume that the Q2 machine has not been obtained and that the number of gems that can be broken per day remains at 20.

http://nhiro.org/titech_hcd/page-67.png

A3: 60% discriminator67 Since there is a 60% chance of getting 20,000 yen, the expected income per unit is 12,000 yen. the purchase price per unit is 9,500 yen, so the revenue per unit is 2,500 yen. since 20 units can be processed per day, the revenue per day is 50,000 yen. the revenue per unit increases by 40,000 yen. Cost 9,500 yenSales 12,000 yenProfit of 2,500 yen per unit

http://nhiro.org/titech_hcd/page-68.png

In other words,68 in this problem set-up, "a device that doubles the processing speed" increases the customer's profit by 10,000 yen, and "a discriminator with 60% accuracy" increases the customer's profit by 40,000 yen. The "identifier with 60% accuracy" has four times the customer value of the "device that doubles the machining speed.

http://nhiro.org/titech_hcd/page-69.png

Summary69 -Customer value is important from a business perspective -Customer value is not "novelty of method" -Customer value does not equal accuracy

http://nhiro.org/titech_hcd/page-70.png

Forced Open Innovation 70 -In 2017, Apple published its first machine learning paper -Even Apple can no longer say no to employee requests to publish -Machine learning technology cannot be a closed monopoly. - "Open innovation is not an option. It is a fact. "**Originally published in 2009 by Secretary of State Rice "Globalization is not a choice but a fact."

http://nhiro.org/titech_hcd/page-71.png

Market Procurement of Expertise71Many companies are looking to market procure expertise, but expert capabilities are difficult to observe and evaluate...(I'm going to tell you a terrible story off the record here.)

http://nhiro.org/titech_hcd/page-72.png

Contract Theory72What is the solution when there is information asymmetry between contracting parties - signaling - screening

http://nhiro.org/titech_hcd/page-73.png

Signaling73 is an action taken by the side that has information. For example, "ability" tries to acquire "signals" that can indicate "ability" there difficult to observe

http://nhiro.org/titech_hcd/page-74.png

Signaling74Why do Apple employees want to publish papers? It is because a track record of publishing papers is a "signal" and an advantage when changing jobs outside of Apple. When sourcing experts in the market, those experts are likely to demand that they can produce a signal.

http://nhiro.org/titech_hcd/page-75.png

Screening75An action taken by the side that does not have the information. For example, offer a "low fixed salary + commission" to salespeople. Salespeople with low ability will avoid it on their own. However...

http://nhiro.org/titech_hcd/page-76.png

Machine Learning and Pay-for-Performance76Machine learning has a characteristic that is incompatible with pay-for-performance: "You don't know if you will get results until you try it on real data. Machine learning experts do not want to take on projects that pay for results. (If the data provided by the business side is dirty, the project will fail to produce results regardless of one's ability.

http://nhiro.org/titech_hcd/page-77.png

We need to find a contract structure that will benefit the smallest start 77 and the most competent professionals. Although a complete solution has not yet been devised, one idea is to make the first step of the contract small and extend it if both parties agree that it will work out well. The experts would not be locked in to a bad project for a long period of time, and the more competent ones would have a better chance of getting an extension.

http://nhiro.org/titech_hcd/page-78.png

Revenue Sharing78 Another strategy is revenue sharing, a method of sharing a fixed percentage of the profit generated by a project with a professional. Both parties cooperate in making a profit. However, it is difficult for the expert to accept a revenue-sharing contract alone because the results will vary depending on the capabilities of the project manager in addition to his or her own skills.

http://nhiro.org/titech_hcd/page-79.png

Summary79 - Forced open innovation - Information asymmetry issues arise in market procurement of specialized skills - Contract theory offers useful suggestions for better contracts

http://nhiro.org/titech_hcd/page-80.png

Overall Summary 80 -Do not make the means an end. -AI does not take away jobs but augments humans = reduced time cost paid by humans -Difference in evaluation axes between academia and business. Customer value ≠ novelty/precision

---