Rise of AI Conference 2018 · Berlin

Hype Meets Reality

Stefan Wess · Empolis

At Rise of AI 2018 in Berlin, Stefan Wess, CEO of Empolis, argues that today's AI hype is only the latest wave of a much older story. He opens with two predictions of human-level machine intelligence, made by Herbert A. Simon in 1965 and Marvin Minsky in 1970, and shows on Google Trends that current interest is still below the 2004 peak. Drawing on his own path from case-based reasoning research to building Empolis, a company spun off from the German Research Center for Artificial Intelligence in 1991 that now serves about 700,000 professional users and 40 million customers, he insists there is no general artificial intelligence today and treats the term mainly as a marketing label for machine learning. He walks through where classic, knowledge-driven AI and newer, data-driven sub-symbolic AI each work well, and names data availability, data quality and bias, and explainability as the three real obstacles to deploying AI in enterprises. He closes by calling for hybrid, "grey box" AI and redefining AI as augmented intelligence that helps humans decide better.

Key takeaways

  • Predictions of human-level machine intelligence date back to Herbert A. Simon in 1965 and Marvin Minsky in 1970.
  • Current AI interest on Google Trends is still below the hype peak reached in 2004, according to Wess.
  • Wess treats artificial intelligence mainly as a marketing term for machine learning, not general intelligence.
  • He identifies data availability, data bias and explainability as the three main obstacles to enterprise AI adoption.
  • He advocates hybrid, "grey box" AI that combines knowledge-driven and data-driven approaches, and calls AI augmented intelligence.

Detailed notes

AI hype in historical perspective

Wess opens by noting that artificial intelligence is part of pop culture and an old dream of humanity, which is why it always makes headlines and triggers strong emotions. He cites two predictions: Herbert A. Simon's 1965 claim that machines would be capable, within twenty years, of doing any work a man can do, and Marvin Minsky's 1970 prediction that within three to eight years a machine would have the general intelligence of an average human being. Looking at Google Trends, he says the current hype level is surprisingly still only about 75 percent of the peak reached in 2004.

From case-based reasoning to Empolis

Wess describes starting to program computer games in the 1980s and wanting, without success, to build a program that could beat him at the board game Risk. He later did a PhD in computer science and worked academically in case-based reasoning, which he explains through the example of a car mechanic who solves a new problem by recalling and adapting a similar case seen before. In 1991 he co-built a company as a spinoff of the German Research Center for Artificial Intelligence based on that idea; today it has about 220 employees and is used by roughly 700,000 professional users serving about 40 million customers worldwide, including, he says, most people in the audience without their knowledge.

"AI is the new smart": from decision support to decision automation

Wess traces a progression in enterprise technology from business intelligence to big data, predictive analytics, machine learning and now artificial intelligence. He states plainly that there is no revolutionary, general artificial intelligence today, calls AI a marketing word that he personally translates to machine learning, and compares it to how everything became "smart" in the last ten years, from smartphones to smart TVs and smart data. He then describes how enterprises currently use data through a decision-support system that produces recommendations a human reviews before acting, and argues AI is moving this toward decision automation, where routine cases run automatically and only exceptions go to humans. Plotting business problems by frequency and complexity, he shows a large middle zone of problems that are neither simple enough to automate today nor complex enough to require full human judgment, and argues AI's opportunity is to shrink that zone.

Classic versus sub-symbolic AI applications

Wess contrasts classic, knowledge-driven (symbolic) AI, which he says has proven business value and reliable, explainable results but is expensive to build and maintain and works best for high-volume, well-defined tasks, with newer, data-driven sub-symbolic approaches (deep learning), which reach superhuman performance on sensory tasks such as voice recognition but are hard to reason about, hard to integrate with existing domain knowledge, and data-hungry.

Three obstacles to enterprise AI

He names three practical barriers to deploying AI in companies. First, data availability: he describes a paradox where consumer-facing systems have abundant data about people but little understanding of them, while enterprises have deep knowledge of their machines but little usable, labeled data. Second, bias: ordinary business data reflects routine cases, while real value lies in the exceptions, so models trained on regular data tend to repeat past behavior rather than create value. Third, and most important in his view, explainability: he asks whether an organization is comfortable with black-box decisions, contrasting an ABS braking system, where no explanation is needed, with a court setting, where liability, accountability and comprehensibility require a decision to be explained even when it was wrong.

Hybrid AI and augmented intelligence

Wess argues the way forward is hybrid AI that combines knowledge-driven representation with data-driven learning, which Gartner calls "grey box" AI between black-box and white-box systems; he places case-based reasoning in this hybrid category. He shows a modified Gartner hype cycle in which he believes the plateau of productivity is closer than commonly expected, describes today's AI as being simultaneously science-fiction general intelligence and a practical engineering tool, and says he dislikes the term artificial intelligence because it invites too much emotion. He concludes that for him AI means augmented intelligence: helping humans make better decisions.

Quotes

Artificial intelligence is a marketing word, and I translated it always to machine learning, because that's where we are at that point in time.

Stefan Wess

Are you happy with black box decisions?

Stefan Wess

AI for me is augmented intelligence, if we talk about decisions helping humans really to make better decisions.

Stefan Wess

Questions answered in this talk

What are the main obstacles to deploying AI in the enterprise?

According to Stefan Wess at Rise of AI 2018, the three main obstacles are data availability, data quality and bias, and explainability, with explainability the most important because black-box decisions are hard to defend in contexts like court proceedings or workers council reviews.

What is case-based reasoning?

According to Stefan Wess at Rise of AI 2018, case-based reasoning is an AI approach in which a system solves a new problem by recalling a similar past case and adapting its solution, much like a car mechanic diagnosing a new problem from experience with similar past repairs.

What is hybrid or 'grey box' AI?

According to Stefan Wess at Rise of AI 2018, hybrid AI combines knowledge-driven, symbolic representation with data-driven machine learning, a combination Gartner calls 'grey box' AI, positioned between fully explainable white-box systems and opaque black-box systems.

Mentioned

case-based reasoning, decision automation, hybrid AI / grey box AI, Gartner hype cycle, explainability, deep learning

Keep exploring

More from Rise of AI Conference 2018

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 2,927 words

Show the full transcript

I would say I will end at the same yeah let's say position as the speaker before but I take a totally different paths so enjoyed right so artificial intelligence is definitely part of our pop culture so we all have in mind r2d2 and Terminator these are the images which come to our mind when we hear that word but it's also part of our common cultural history so every civilization on earth had that idea of creating artificial beings so it's a kind of an old dream really of humanity and that's the reason why artificial intelligence always has the headlines that's why we get really emotionally connected to that topic and have yet to headlines one if machines will be capable present twenty years of doing any work a man can do just the first one and the second one in from three to eight years we will have a machine with the general intelligence of an average human being any idea from what year that headline is so surprise surprise the first one is from 1965

from H I Simon and the second one is from Marvin Minsky in 1970 so being there done that got the t-shirt so if you do some analytics and just look on Google Trends I really was surprised when I saw that the hype in 2004 and that was the farthest way I could go we are currently at 75 percent of the level we had in 2004 so how did I get involved in that ai gang in 80's I started programming computer games like like defender and we also play the board game at that time which was called risk in German physical it's a game about world domination and my personal dream was I wanted to create a computer program who was able to beat me in that game it was a total disaster I was not able to do that and it went beyond my programming skill and beyond any technology which has existed in that point in time so I did my PhD in computer science I put the score here from ResearchGate which is

a start-up here in Berlin and to give to talk a little bit more on credibility to compare the numbers I put here the numbers from ashen Ram who is the leader of Google AI and of course he has better numbers yet and I because he works for Google but it was not the point we both worked on an academic field which was called case based reasoning which was or is in the eye paradigm in building our eye system and what is cases reasoning about just in two minutes so you think about you have a problem with your car it's a new problem and you have no idea what you need to do so you go to the guy with just a repair guy what is the main difference between you and that guy of course he has more experience he has seen more cases he has more knowledge about that so what does the repair guy doc then do ok he has seen a similar situation and there is that car who seemed to have

a similar problem he applies that kind of solution to your problem and of course he has learned something problem solves and the world and your car is getting a little bit better so that's the idea of of case based reasoning and we build out the company out of that idea in 1991 as a spinoff the German Research Center of artificial intelligence so as of today we have about 220 employees the 500 company is out there using it about 700,000 professional users and we serve about 40 million customers worldwide and my bet would be that everyone in the audience had been in touch with our systems there without knowing it so and then something happens so we all had that picture in mind of beating best player and go so it a really big advantages in deep learning and deep learning is a part of artificial intelligence it's machine learning and the reason why we had that kind of success is because there's a lot of data we have the hardware and as mentioned before all

the software's open-source so we got a long way to be there so we started with business intelligence that big data as a high prima analytics machine learning and now we are at artificial intelligence and just to make that very clear we are very well there's a topics above but as of today standing here on the stage there is no revolutionary artificial intelligence if we think about like general artificial intelligence but being in that field for a long time my complete surrender so artificial intelligence is a marketing word and I translated always to machine learning because that's where we are at that point in time and for me artificial intelligence is the new smart yeah with the smart phone the Smart TV the Smart Watch the Smart Data the smart services so everything really got smart in the last 10 years and I think the same is happening here was artificial intelligence so in the next slide I will use artificial intelligence as a term but you can always translate it to machine learning so what

is the reason why a lot of people and why there really is the hype in the enterprise why there are so many people interested in that kind of technology and the main reason is that there's so much data out there I want to make sense out of the data and finally we'd like to make smarter and better decisions so that was this report the saying so what does this mean as of today in the enterprise we have a lot of data and finally you're going to have actions so in between that's a system and this is here called decision support to analyze the data and make some recommendations then usually we have a human who analyzes these kind of recommendations the human makes a plan then he puts his ideas in another system and finally the other action is performed so now we are going to the area of decision automation which means that the regular business procedures will fly through and only for the exceptions we need the humans what does this mean so

if we look on the problems in the enterprise and we have the frequency and the complexity of the problems in the upper left corner we have problems which require very frequent purchases are really easy down on the right corner we have problems which result in but are very complex as of today in the blue area the easy problems we can automate by software hard problems will be solved by humans and in between there's a big white spot what do we do in that area and as of today mainly human labor is performing these kind of tasks so if you look on the problems and how often they occur you have these kind of common problems which are very often hard problems and something in between which other sporadic problems the common problems easy cost we can do that really fast hard problems really complex we need a lot of work and the sporadic problems just some are in the middle if we multiply that and this is the workload in an adverse organization so we

see for the common problems there is little chunk how problems stay and there are a lot of problems in the middle and from my point of view but it AI is able to do it can shrink down this blue area so that if you finally if you can treat that area and reduce the effort so this is supported by your pen you study here from pack which is called what can a I bring there to business applications and finally our goal is to do something like this so have a high automation so that the white area is spotted out and run by machines so if you look at the classic AI applications they have a proven high business value they have some advantages so they provide really good results they are very compatible but there are very expensive in trading and in maintenance so we have a kind of very restrictive business application areas so mainly high-volume tasks the simple reasoning tasks and always a well-defined business scenario is what is the most unsuccessful application

in that area in my view business navigation system you might not think that this is a nice solution but finally it's based on the principle of the res a deep blue is beating Kasparov in chess so the new application which other sub symbolic wants have also some advantages so they have these kinds of superhuman capabilities but the disadvantage says our reasoning is really difficult and there's no integration you have this existing domain knowledge so there's also proven high business value yeah but also a kind of restricted application areas you need a lot of massive I need a lot of data you're somehow restricted to these kinds of sensorial tasks so when finally is the kind of always this kind of stimulus and response areas most successful application is the voice recognition assistance we know as of today so if we talk about applying these technologies in the enterprise what are the killers from my experience in really deploying these kind of technologies and from my point of view they are free and you might be

surprised by the last one so the first one is really data availability the second one is data quality and the last one is explained ability and comprehensible 'ti let's go shortly in these topics the first question is how to obtain these massive data sets when we go to an enterprise a lot of people say okay there's so much data out there why we just not use it the reality in most companies as of today is there is not that kind of data we need to have for these approaches and they're something which I call a paradox on because if we talk about humans then we have a data rich and the knowledge poor environment so which means we have a lot of data about all of you but we don't understand why you are doing that so it really there's the data but not the knowledge if you go to the enterprise and this is fascinating but it's the opposite so we have a lot of knowledge why your machine is working because we have

built that machine but not a lot of data are really labeled data we can use so yet data acquisition is really the first killer the second killer is the bias in the algorithms and the data so and I don't talk about these ethical questions like Microsoft a in women and man so even if we go you have you bring it down and we really talk about business there's the same problem and that's because you have a lot of data about your regular business which are the Penguins they all look similar but the real value lies in the exception which is the red apple in the middle of all these green apples and if you just let your regular business students go is really really difficult to find that data about really that value part of your business so you always tend to have the bias do like as you always done before which is not really creating value but the most important the killer question and I'm sure you didn't expect that is the question

are you happy with black box decisions because on the one side all these algorithms work like input-output stimulus-response so can you live with just a result without any explanation I can if you think about your braking system a pH HBS in the car so you just want your Cobb stopping as fast as it can and you don't want to know what's going on so mote manager says I want to go for the results and I don't care how we achieve these kind of results the reality is that works really fine before you really get in front of court and you get in front of court and then you need to explain your decisions and there is really a big difference between how these systems work and how we work really with our law and this is not just one law it's the question or the principle how we as humans have created the law so there are questions like liability accountability responsibility comprehensibility if you want to talk to the workers council and there's always

the question every on front of court explained me your decision even if it's the wrong decision tell me why you did that that way even it was wrong I need to have an explanation so the final and from my point of view most important question is are you happy with black box decisions and can you defeat that so this is the question of of results really versus law so we have on one side this old and classic sub symbolic are algorithms or symbolic algorithms and two new trendy your sub symbolic algorithms on the left side these approaches are data-driven on the right on the left side they are knowledge driven on the right side yeah there are data-driven and what's really interesting is if you look at the costs of a real deployment of these kind of applications so because of your knowledge driven you need to put a lot of manual work it's really expensive to build these kind of applications because you need these kind of authors and experts if you are data-driven

you're the same costs apply but it's not for creating all these kind of knowledge but providing all the data and providing the data quality really which is needed so in my point of view moving forward we need to have something like these kind of hybrid AI approaches which for example do not fall under this problem was explaining ability so you had on the one side the kind of representation and on the other side the side of learning and you need to combine these two and coming back to case based reasoning the kissters reasoning is one of these approaches which is hybrid a gardener calls this grey box AI so combination black box white box gray box so if really data is the fuel of the 21st century which I truly believe then hey I somehow the engine who makes this fuel really powerful and I put the Gartner hype cycle on the slide we all know this hype cycle in my personal view we are moving up on the peak of inflated expectations but I

have photoshopped it a little bit because it will not go as deep as expected and the plateau of productivity is really really close so in my point of view artificial intelligence as of today is like Schrodinger's cat so on one side it is really that on the sides that we think we have these super general artificial intelligence who's able to solve every problem we can imagine so this side is that but it's alive if you think about an engineering solution which helps us to improve our lives and solves we get really big problems of humanity and just for the geeks and I don't we'll go into that if there any computer scientists out there so they understand so even for artificial intelligence the famous question of computer science is still is still unsolved I personally never liked career constellation until against and I said before I hate artificial intelligence as a name because it creates so much emotions and if we are always in science fiction mode we are not able to see the reality

exploit the reality and also work with the problems of reality and that I think is really important because there are so many advantages of AI as the technology that we need to be careful not overdoing that and that's why I don't like to work Keo AI and I have my own names for this so for Korean and this I'm doing for last 10 years Korea's future computer science because as seen with navigation systems you might think this guy's crazy on the stage but reality in 2000 this was artificial intelligence that was the edge of technology in 2097 in 1970 blue has beaten Kasparov 2000 navigation systems so completely informatics is ki and I said I'm ending at the same as my really famous speaker before AI for me is augmented intelligence if we talk about decisions helping humans really to make better decisions having intelligence augmented all over and that directly goes to division yet the speaker before has so thank you very much I'm perfect in time and if you are interested in that

topic we provide the monthly AI newsletter you can go to our home page and subscribe to it and I'm happy if you follow me on Twitter on LinkedIn because they're always I post your stuff you might find interesting or not so thank you very much

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.