What does artificial intelligence mean for mankind's future?

A lot of people talk about artificial intelligence, but few are actually knowledgeable about the topic. To learn more about this interesting field of research, our e-learning expert KTU sat down with Dr. Andreas Müller.

There has been quite a fuss recently over the poker software PokerSnowie, which functions with a self-learning neural network. Given a new Terminator movie has just been released as well, one could let the mind drift a little and wonder if artificial intelligence could actually pose a threat to mankind!

Dr. Andreas Müller is a qualified mathematician, and has a doctorate in the Computer Science field of "Machine Learning". He's currently working at the University of New York on the project "Scikit-learn". Before his return to university, he also worked for Amazon as a "Machine Learning Scientist".

Dr. Andreas Müller

KTU: Hi Andy! Thanks a lot for taking the time to shed some light on the topic of artificial intelligence for our members. Your focus of research is called "Machine Learning". Is this supposed to be differentiated from artificial intelligence or does it mean the same thing?

Dr. Andreas Müller: In principle, artificial intelligence means to create a machine or computer that is able to act "intelligently". However, "intelligent" is quite a broad and ambiguous term. Artificial intelligence as a research field mostly deals with logical conclusions and the search in familiar and limited systems. A good example for that is the game tic-tac-toe or, searching for the shortest possible route through a maze.

Implementations like that usually have very clear rules and the challenge is to act in a complex but well-understood system. Artificial intelligence was a very common term and field of research in the 60s and then again in the 80s. Unfortunately, researchers had to realise that the real world doesn't look like a tic-tac-toe field at all and instead contains ambiguities and a complexity that algorithms for artificial intelligence cannot handle. Therefore, research reoriented towards algorithms that are trying to learn through statistical processes instead of modelling the entire system beforehand. These algorithms are usually summarised under the term "Machine Learning" and most implementations that contain "intelligent" components now use these kind of statistical methods.

Facial recognition through "Machine Learning"

A good example for a "Machine Learning" algorithm is the facial recognition that's part of any smartphone camera these days. The developers didn't determine what a face should look like, but instead wrote a program capable of recognising and distinguishing specific patterns in a picture. They then ran a large database of pictures with marked faces through that program, enabling it to learn what's distinctive about faces and how to recognise them.

KTU: What are you working on at the moment, provided a normal person can vaguely understand what it is about?

Dr. Andreas Müller: Currently, I'm mostly working on the development of Scikit-Learn. This software enables developers to use "Machine Learning" algorithms in their own programs. Basically, it's a toolbox that researchers and developers can use in order to allow learning from data. It's used by a lot of big companies like Facebook, Amazon, the New York Times, OKCupid and Twitter.

KTU: From your point of view, what are the most impressive developments in artificial intelligence or "Machine Learning" within the last 10 years?

Dr. Andreas Müller: A breakthrough in "Machine Learning" has been the broad implementation of neural networks, often referred to as "Deep Learning", especially with regards to image and text data. This technology has actually existed since the 80s, but only in recent years have we been able to reach enough computing power and get enough data to implement neural networks effectively.

Neural networks for automatic translations

Two particularly interesting implementations are image recognition and automatic translation. In image recognition, algorithms based on neural networks can already precisely recognise thousands of different object categories. With regards to automatic translation, neural networks can translate text between two different languages with a high accuracy. The really impressive part of this is that neural networks can solely learn with the aid of sentence pairs. For example, a large amount of English sentences and the respective Chinese translation is enough to create a general translation program. Skype is currently working on implementing an automatic translator capable of directly translating what's been said. A very interesting article on this topic is: The Unreasonable Effectiveness of Recurrent Neural Networks.

Another remarkable implementation is learning to play Atari: DeepMind’s AI is an Atari gaming pro now. The outstanding part of this work is that the researchers of DeepMind were able to create an algorithm that's at least as good as and often better than humans in roughly 50 different games. Usually, gaming algorithms like that are only geared towards a single game.

KTU: There is quite an unsettling video by cp grey which basically claims that computers will be a cause of unemployment for large parts of the population within 5 to 10 years. Do you share a similarly drastic outlook? Which jobs are secure in the sense that they can't be performed by computers within the next 10 or 20 years? Which jobs will still require humans in 150 years?

Dr. Andreas Müller: I don't think that a large part of the population will become unemployed due to automation within the next 5 to 10 years. Existing trends like self check-ins at airports or self check-outs in supermarkets will continue to spread. However, these two applications don't require a lot of artificial intelligence. cp grey's claim is mostly based on "General Purpose Robotics", meaning machines that can perform a lot of different tasks. I'm convinced that he's off the track with that assumption, especially regarding the next 50 years. I think we will see a lot more specialised machines before humanoid or "General Purpose Robotics" become feasible. A really interesting specialised robot that we might see in the near future will be cars. Maybe taxi drivers will be replaced in 10 years, maybe 20 —however, definitely not by a humanoid robot sitting in front of the wheel, by an on-board computer.

Apart from that, I think most jobs are secure for the next 10 to 20 years. Of course, some customer interaction will certainly be automated, but there will still be enough work in the service sector. For many companies, automation simply means that they will be able to scale. It will allow them to achieve more with the employees that they have and the goal won't be to replace them completely. What the world will look like in 150 years is really hard to predict, but service and communication are sectors which humans will still be working in for a long time.

KTU: "Technological singularity" would occur once an AI is "clever" enough to improve its own intelligence and set its own goals. Will we ever witness that day?

Dr. Andreas Müller: Singularity is a bit of weird term and implies blurring boundaries between solution of problems and consciousness. I think we will witness the day when it will be very hard to find a tuning test that a machine can't pass. In order to improve itself, the AI would have to be capable of programming a better algorithm. I wouldn't exclude that from happening as well. However, I'm not sure if I would actually call that singularity.

KTU: Let's assume a super-intelligent robot like that exists. Is it possible to make any assumptions about the type of goals it would set itself?

Dr. Andreas Müller: Intelligence and robotics actually don't have a lot in common. In a lot of science-fiction visions robotics and artificial intelligence go hand-in-hand. However, in research, both fields are pretty independent from each other. In robotics, machine learning is used to detect objects, grab something and recognise people. However, research in machine learning is almost never connected with robotics. There are researchers arguing that interaction with the world is critical in order to learn something about it. But that is not really mainstream research. Regarding the type of goals such an AI would set itself: that's not a concept that's currently part of any research I'm aware of. Usually, you have a specific goal in mind that you want to reach with an algorithm. What would it mean to leave the objective to the algorithm itself?

Poker is not a game with "full information"

KTU: Some weeks ago there was a big showdown between humans and machines in poker. The humans were able to come out on top. What makes poker that much harder to master for a computer in comparison to chess?

Dr. Andreas Müller: You know more about poker and chess than I do. There are some clear distinctions: chess is a game with "complete information". Both players always know what's currently happening. In poker, there are a lot of different possibilities with regards to the hand the other player could be holding. When a computer makes a decision, it basically has to take into account all these different possibilities. That's what makes poker that much harder. Additionally, I think the scholarly interest in poker is much smaller than it was for chess. Before "Deep Blue" many people were convinced that a computer can't be "creative" enough to win a game against a human. Now, it's relatively clear that it's only a matter of time until humans are beaten in a specific game. Therefore, it's not that interesting to research in this field.

KTU: Do you think an affordable robot will be available at some point that can take over most of the daily routine and follow commands like "Please fetch little Kevin from school", "Clean up the living room" or "Please prepare a meal for 7pm and iron my white shirts"?

Dr. Andreas Müller: I was working in a research group that was trying to develop household robots during my PhD. From an "intelligence" perspective, we will soon be able to understand such commands and even do the necessary planning for some of them. It's relatively clear what needs to be done to fetch Kevin from school or iron a shirt. Cleaning up the entire living room could be a bit harder, but not impossible at all. While machine learning has made big steps forward these last years, I'm not sure if mechanics are able to catch up. I'm not a specialist in robotics, but most robots that I've seen so far are very slow and often imprecise. The only exceptions are the robots by Boston Dynamics that are operated by hydraulics.

These robots are amazingly fast and strong but better suited for military use because of the heavy and noisy technology. You don't really want to put something like that into your living room! The best robot for the job would currently be PR2, a research robot worth $400,000 that's able to fold a towel in a few minutes thanks to its precise manipulators and sensors. Still, we are quite far from a feasible solution.

KTU: If someone is interested in artificial intelligence or machine learning, is there a book or a website that you could recommend to get started?

Dr. Andreas Müller: The standard publication is "The Elements of Statistical Learning" which is totally free and available for download.

KTU: Thanks a lot for this interview, Andy!

Arne "KTU" Gewert

Arne KTU Gewert is Senior E-Learning Expert at PokerStrategy.com in Gibraltar and responsible for a large portion of our offerings.

    Comments (3)

    newest first
    • TraFilip

      #1

      Nice article
    • ZeroDegrees

      #2

      That was one of the best articles I've read here. Very interesting and informative. Thumbs up!
    • IvicaIliev77

      #3

      Loved it. More articles like this, future is in AI for sure.