In 1936, Alan Turing presented a seminal paper to the London Mathematical Society titled, “On Computable Numbers, with an Application to the Entscheidungsproblem” (one of those compound German words which means ‘decision problem’) . Building on the work of earlier mathematicians and logicians, Turing’s thesis sets out proofs for a hypothetical machine which would be capable of performing any mathematical computation if it were programmed by a string of instructions — an algorithm — on a memory tape. This “universal computing machine” would become the eponymous grandfather of the modern-day computer, and would be instrumental in paving the way for artificial intelligence (AI). Seven years later, my father, Hock Lin, was born in a soon-to-be independent colony on the fringes of the British Empire, called Singapore.
After the War ended, and when my dad was still a toddler, Alan Turing was finally realising his prototype for the commercial market by consulting and programming software for Ferranti, a British electrical engineering company . The Ferranti Mark 1, the world’s first stored-programme computer, was deployed in 1951. Shell became one of Ferranti’s early clients, and purchased several of these large machines for their laboratories in Amsterdam . By the late 1960s, when my father was entering the workforce as a young chemical engineer, computers were already well integrated into the global operations of the oil and gas industry.
My dad left the industry in the 80s to become a pastor, and today he uses a MacBook Air for his preaching and teaching with some proficiency. So, although he’s been working with computers for five whole decades, I thought it would be refreshing to catch him up with how machines are learning today. We sat down at a cafe in LASALLE College of the Arts, where artist, technologist and educator Andreas Schlegel heads up the interdisciplinary Media Lab, funnelling students from all of the college’s faculties through an open-ended incubator for them to experiment with technology to create all sorts of artefacts and prototypes.
CJC: Dad, can you tell us about the first time you encountered a computer?
HLC: My background is in engineering, so during my time, linear programming was just coming into the engineering world. I remember going down to IBM at Raffles Place in the former Shell building with all our punch cards.
CJC: Which year was that?
HLC: It was back in the 60s… 1967 or 1968.
AS: I recently saw an old Zuse machine , a vacuum-tube computer with punch tape reader from the 1950s, in an art museum in Germany.
HLC: A museum!? [Laughs] At that time, I was working for Mobil Oil for their refinery in Jurong. IBM had a computer that was operated with punch cards, disc and tape drives. We would carry one or two boxes of the punch cards with reels of magnetic tape containing the software that we used to run our programmes. The punch cards fed the data that were the input to optimise the refinery operations. The data included variables related to the type of crude oil (feed properties), feed flow rates, as well as product properties and flow rates. Limitations included capacities of the different units of the refinery. A typical programme would have 200 to 300 variables and a similar number of constraints. Costs of crude oil was a major constraint, as were the prices we could sell the different products for. The optimum results we looked for typically were the mix of products (fuel oil, gas oil, kerosene, gasoline, and so on) we could get from the crude oil feed within the limitations of the different units of the refinery complex. It would be a long process of iterations using linear programming.
AS: That’s quite ahead! It sounds like a lot of “if/else” conditions.
CJC: How long would it take to run the programme?
HLC: Well, we would usually be there for about an hour.
AS: Probably a big obstacle back then was that you couldn’t see immediately what you were programming, so you needed a trained imagination to predict what those machines would spit out at IBM. And only they were able to process your programmes and those boxes full of punch cards.
HLC: Yes, that’s right. When did artificial intelligence come into the picture?
AS: Despite the current hype about AI, it has been around for some time. In the 1950s, scientists began experimenting with artificial brains and neural networks called perceptrons. But the depth of AI-related research then went up and down over the years. Some results were promising, others less. I guess it was often difficult to maintain the interest and funding of AI research, as there were no obvious or limited benefits. As you mentioned earlier, linear programming was more effective and the focus was largely in this direction.
After a kind of AI “winter” in the 1970s, promising developments were made in the 80s by rediscovering back-propagation to efficiently train artificial neural networks. However, AI still remained within scientific and academic communities since there was no obvious market value yet.
In the mid-2000s, a technique called deep learning gained popularity and presented some very impressive results. What drew my attention was a project called DeepDream , a computer vision program that enhanced patterns in images, and produced very hallucinatory-looking and fascinating images mixed with dog or snail-like creatures.
From there, deep learning moved very quickly from the academic and scientific community to the business side of applications. In summary, it can be said that AI has been around for quite some time, but now with deep learning, AI as an umbrella term has entered our minds and has caused a lot of attention.
HLC: To clarify, is deep learning the same thing as artificial intelligence?
AS: It is a subset of machine learning in AI.
We should also distinguish that there is a general artificial intelligence that people seem to be very afraid of because it tries to make the machine very human-like. Then there is narrow artificial intelligence that makes computer programs really work very effectively within a certain domain to solve a specific problem by performing a single task. However, if the context of the problem does not fall within the scope of that domain, that system will most likely fail. It only makes really good predictions within a limited context.
HLC: I’ve seen something like that in a TED Talk by a robotics researcher, Peter Haas . He presented a picture of a wolf-looking animal, and he asked a question, “Is this a wolf or a dog?” The AI evaluated it as a wolf, but it was actually a dog. He said, “When you look at the picture, what do you see that could help you discern what kind of animal it is?” For me, I thought it would be the eyes, or maybe the ears — how they stand up — or maybe the teeth. But actually, the AI wasn’t like that. It profiled it on the basis of the background, where there was a lot of snow. According to the input, wolves are connected to scenes of the North, where there’s lots of snow. But this was incorrect.
CJC: There are many occasions when the machine can be mistaken, because of some unseen bias inherent in the data.
AS: A shortcoming in the early uses of artificial intelligence and machine learning was that there was not enough data to work with and train with. Deep learning applications today are trained on very large data sets — the more data, the more accurate the predictions can be. With social media and the Internet, we share a lot of our data, intentionally or unintentionally: images, texts, clicks, conversations, videos and, as a result, machines have access to huge pools of data, labels, information and relationships that can be used for training.
CJC: What would you say is the minimum amount of data for a machine to learn effectively?
AS: I think it depends a lot on the media used. If you take a set of pictures and an image is about one megabyte in size and a million pictures are used for training, you get at least one terabyte of data. The more data you can provide, the more detailed the data set, the better the model, results and predictions. These early data sets were quite small, so the result was limited.
Deep learning applications today are trained on very large data sets — the more data, the more accurate the predictions can be.
HLC: Do you know how much data Google has got?
AS: I wouldn’t know the number. My dad always likes to use “peta”. We have giga, we have terabytes, then there’s petabytes. I think that’s 15 zeroes or 10 to the power of 15. It’s just a gigantic number, but even peta is by far not enough to describe the amount of circulated and stored data available today.
But data alone will not do the job. To process all this data, you need a lot of computing power. In addition, you need to understand this data, like through labelling, which is often done by human work crowdsourced through applications like Amazon’s Mechanical Turk. In this case, workers are paid to look at pictures and describe their characteristics by being asked questions like, “Is it a dog or isn’t it a dog?” Images are then labelled with a positive or a negative answer. So it is not only about labelling what a thing is, but also about what it is not.
CJC: It’s the same with Google or Facebook, where they outsource to smaller companies and platforms in India, Bangladesh and the Philippines. Crowdsourced, cheap human labour trains the AI. How exactly does this happen?
AS: It’s a simple but repetitive task. Based on a series or a pair of pictures for example, workers click on the one closest to the answer to the question asked. They earn per click and this can go on for hours and hours.
HLC: But at some point, the machine takes over, right?
AS: I would say so, yes, humans do prepare these datasets. After the machine has been trained, it almost perfectly predicts an output based on an untested input.
How does deep learning work? Let me try to explain. It is a combination of a pre-trained model, often based on a large data set, and a neural network. The network consists of many small processing units, called neurons. Individual neurons are connected to each other and transmit signals to each other when activated. These artificial neurons receive inputs from which they can generate an output. These outputs are then forwarded to the next neuron. At each layer of the neural network, different tasks are performed on the data passing through. In the case of face recognition, for example, features from the pixels of an input image are extracted across multiple layers on the network to then determine whether or not a face is contained. It can also be predicted where a face is located in the image, and possibly additional information like age, gender, and so on can be extracted. This procedure can be performed on faces, cars, objects, speech, sounds, texts and many other things.
HLC: Could I say that for AI to excel, it has to depend on the amount of data that’s available, as well as the computational power to process all that data, and the availability of humans to train the data?
AS: Exactly. Well, that’s what I thought too, but scientist David Silver from AlphaGo argues  that it’s not so much about the amount of data or computation anymore, it’s more the algorithm — how machines learn for themselves through reinforcement learning, and this is what we are seeing more often. This is learning through trial and error, rather than through a big data set approach. In a reinforcement learning environment, we have an agent which performs an action. This action has an effect on the environment, and based on the effect the agent gets rewarded or it doesn’t. Over time, the agent learns which of the actions performed previously result in positive or negative outcomes, and on that basis, prepares for the next action to be taken to finally complete the task. Simply put, the system begins quite stupidly with random actions, but over time, through iteration, it learns to optimise and perfect itself. This is what AlphaGo and AlphaZero have been accomplishing lately.
HLC: And that’s where iteration comes in. With each iteration, there is a refinement.
AS: Yes. I recently read in a book called AI Superpowers by Kai-Fu Lee  how China and Silicon Valley are competing with each other technologically. The author mentions four waves of AI that I found very relevant to our discussion.
The first wave is “Internet AI”, which we already know, and as we surf the web, we get recommendations for the products we should buy at Amazon, our Facebook news feeds are personalised, and so are our Google searches. All of this is due to machine learning algorithms that optimise the results of the information we provide.
The next wave is “Business AI”, and the extent to which insurance and banking systems, stock markets or other sectors such as health, housing, gaming, etc. benefit from these algorithms to optimise their results and revenues.
The third is “Perception AI”, which depends heavily on computer vision, object recognition or face recognition to create intelligent environments or even smart cities. I think that Perception AI is something we don’t really see or recognise, but it’s partly here and already makes certain parts of our lives easier, like infrastructure or traffic management.
And the fourth wave is “Autonomous AI”, where we expect machines that can work independently, like self-driving cars, robots and so on. This is probably closer to what we were talking about earlier, a general artificial intelligence. Up to this point, we do live in a world surrounded by narrow artificial intelligence.
HLC: I’ve got a question. Can machines come to a place where they can make moral decisions?
AS: I don’t know. I think that also depends on how they are trained and who trains them. At the beginning of our discussion you spoke of a bias of the AI that identifies a wolf only based on the background. Bias is a big issue. Only certain expert groups train these machines, and applying an objective and neutral world view can then become a challenge.
HLC: Then the real question is, who determines what is right and wrong for those machines? Other than that, it’s just amazing what these machines can do. Frankly, that’s my main concern, because this is my profession now — helping people make moral choices so that we can have a better world.
Can machines come to a place where they can make moral decisions?
AS: But would you want machines to learn morality by themselves so they can make better choices than humans? Do you think we humans are a good example for machines, because we’re already so…
HLC: Messed up? Well, you can’t judge a system by its abuses, even though that’s what we tend to do. But no, you cannot do that, you see. You’ve got to judge the system on the basis of the principles that it’s built on. And principles are based on truth. One basic, fundamental truth is the dignity and worth of every human individual. If we don’t agree on that, then it’s just the survival of the fittest.
AS: That’s quite a big undertaking, no?
HLC: Mao Zedong said, “Power comes from the barrel of a gun.”  Is that going to be our definition? Whereas, when we respect the dignity of every human individual, that’s when we have the best chance of having a civilised world.
AS: Do you think machines, at this point in time, can help?
HLC: I’m not sure. Like you said, a big part of it is who is programming the machines.
AS: Efforts are being made to challenge the biases in programming, in data and in algorithms. But still, Silicon Valley is a largely male-dominated, data-harvesting, and profit-driven environment that defines most of the technologies we live with today and possibly in the near future. Also, I often wonder whether it is a good thing that technologies and their products are predominantly motivated by scientific, engineering and entrepreneurial ideas? I tend to not think so. Then how do people like me, artists working with technology and you, as a journalist, and you, as a moral teacher, and others — how do we give ourselves the opportunity to contribute ideas and participate technically and otherwise in this development?
I often wonder whether it is a good thing that technologies and their products are predominantly motivated by scientific, engineering and entrepreneurial ideas?
HLC: Yes — to enter the conversation.
AS: And I think by doing so, we will also have a more diverse range of machines.
Turing, A.M. (1937) [Delivered to the Society November 1936]. "On Computable Numbers, with an Application to the Entscheidungsproblem". Proceedings of the London Mathematical Society. 2. 42. pp. 230–65. doi:10.1112/plms/s2-42.1.230.
Swinton, Jonathan (2019). Alan Turing's Manchester. Manchester: Infang Publishing.
Erno Eskens; Wessel Zweers; Onno Zweers. "Interview with Lidy Zweers-De Ronde, programmer of the MIRACLE (Ferranti Mark I*), the first commercial electronic computer being employed in the Netherlands at Shell labs in Amsterdam". https://onnoz.home.xs4all.nl/miracle/extra/texts/InterviewwithLidydeRonde,Miracleprogrammer.txt Retrieved 1 September 2019.
The Z22 was another early commercial computer developed by German computer scientist Konrad Zuse in 1955, shortly after the Ferranti Mark 1.
DeepDream is a programme created by Google engineer Alexander Mordvintsev which uses neural networks to enhance patterns in images, creating dreamy, hallucinogenic results. In 2015, Google posted the computer vision software on GitHub, and many more applications sprung up so that the DeepDream technique went viral. Check out the original open source here: https://github.com/google/deepdream
Watch the talk here: https://www.ted.com/talks/peter_haas_the_real_reason_to_be_afraid_of_artificial_intelligence?language=en
Founded in 2010, DeepMind Technologies has created a neural network capable of playing video games and board games similar or even better than humans. Google’s parent company, Alphabet, acquired DeepMind in 2014. Watch one of their lead scientists, David Silver, discuss the company’s latest discoveries: https://www.youtube.com/watch?v=tXlM99xPQC8
Lee, Kai-Fu (2018). AI Superpowers: China, Silicon Valley, and the New World Order. Boston: Houghton Mifflin Harcourt.
The full quote is: “Every Communist must grasp the truth, ‘Political power grows out of the barrel of a gun.’” Zedong, Mao “Problems of War and Strategy" (November 6, 1938), Selected Works, Vol. II, p. 224.