How machines learnWhen I was a member of the faculty at IIT Kharagpur, one of my senior colleagues had told me that we should not to teach students but instead, help them learn. Can this idea to extended to machines?
Any spreadsheet user would know how to have a computer add two numbers. Why just add? You could do many other tasks like finding percentage or net present value as well. Obviously someone has written a computer program that does all this! Those who know computer programming would also know that writing the code to add two numbers is quite easy. Depending on the level at which you would like to describe the process, the program can be written in a high level language like Python or Java, with just one line of code. But if one uses a low level language like Assembler or binary code then he will need a large number of rather arcane instructions written as a series of 1s and 0s [http://bit.ly/add2numbers]. Irrespective of the programming language used, a program that performs any task on a computer, from adding numbers to showing a YouTube movie, is a series of explicit instructions given by a human programmer. Irrespective of the complexity of the task, it is always a human who teaches a computer how to perform it. This is one of the fundamental tenets of computer programming … or was, until the emergence of machine learning. Instead of teaching a computer, can we make it learn on its own?
There are many facts that a child is taught in school -- like how to add two numbers or when Raja Harshvardhan ruled in Kannauj. But there is much more that he learns on his own -- like recognising his mother or realising that toffees are good to eat but stones are not. How people learn, for example, to recognise a friend in crowd or choose the best move in a game of chess, is something that has baffled computer scientists for a long time. That is why even though we have had computer programs doing incredibly difficult things -- like landing a spacecraft on Mars -- they have had immense difficulty in performing apparently simple tasks like crossing a busy city road. This has now triggered a new line of thought that, instead of teaching computers how to perform, we need to equip computers with the ability to learn how to perform!
While a lot of human learning involves memorising facts, the really complex or intelligent skills that people acquire are based on a series of trials and errors that they make as soon as they become aware of their environment. Whether babies or laboratory rats, intelligence is acquired by performing a task and determining whether the outcome was good or bad -- was there a reward or a punishment?. This conscious feedback loop spread over days or even years gives man the ability to do intelligent tasks like recognising a friend in a crowd even if he has grown a beard or lost his hair. After despairing for years about what kind of instructions that must be given to a computer to do similar tasks, scientists have looked into the human brain -- the biological marvel that sits inside the cranium -- for ideas and have finally hit upon a way to address these challenges.
The animal brain is an electrochemical device that consists of, literally, trillions of cells called neurons that can sense, generate and transmit electrical signals. Each neuron can be visualised as a little blob with a number of wires sticking out of it. Most of these wires, called dendrites, can sense an electrical symbol, while one, called the axon, can generate an electrical signal depending on the signals that it has received on its dendrites. The signal generated on the axon of one neuron feeds into the dendrites of other neighbouring or distant neurons, creating a complex electrical circuit that is constantly sending electrical charges rushing around the brain. Neurologists have a rough idea of what kind of electrical activity is associated with the corresponding human behaviour and it is believed that memory is a created or defined by the way these trillions of neurons are connected to and influence each other.
Computer scientists have merged these two mechanisms -- the reward-and-punishment mechanism from behavioural science and the input-output electrical mechanism from neurology -- to create a software program that mimics the behaviour of biological neurons. Such a software program is called an artificial neural network (ANN), and is the basis for almost all intelligent software including AlphaGo -- that can beat humans in board games, Watson -- that can diagnose clinical diseases, or the Google Self Driving car -- that has travelled thousands of miles with an accident rate that is far less than that of humans.
While we already know that a computer can be taught to add two numbers, let us now see how it can taught to learn to add on its own! To build a general purpose ANN one starts by simulating nodes, or artificial neurons, each with its own set of inputs and an output. Each input is associated with a parameter, or number, called weight and each output depends on the inputs and another parameter called bias. Output from each node is sent to other nodes to mimic electrical signals on the biological neural network. The number of nodes along with the numerical values of the respective parameters define the ANN. When the ANN receives an input it generates an output that depends on the specific values of weight and bias parameters. If the values are chosen at random, the output is wrong but with the right values, the network will generate the correct answer. Learning consists in determining the right values. The beauty of this approach is that almost all problems can be addressed with this structure -- only the number of nodes, weights and biases need to change.
To teach a computer to learn how to add we provide it with training data, consisting of thousands of addition problems along with their correct answer! Initially, the ANN has random values of parameters and as it checks with the given correct answer and determines the error, it keeps changing the parameter values until, after thousands of attempts, the error becomes very small. Thus the machine discovers, or arrives at, the correct parameter and can be said to have learnt addition. Now it can solve new addition problems correctly even though no human programmer gave it explicit instructions to add or even the right parameter values. The programmer’s skill lies in specifying how the parameters should be changed so that the error is minimised with the least number of trials. What is amazing is that a similar way of changing parameters can help the machine solve many other very different problems.
This is meta learning. The machine has learnt to learn and can now address many other hard problems like handwriting recognition, face recognition, detecting criminal behaviour like financial fraud and computer hacking, playing chess and other board games, medical diagnosis, weather forecasting, road navigation and even reading human thoughts! In all such cases, the program has to cycle through thousands of pre-solved problems until it discovers the values of the parameter that makes it generate correct answers. Sometimes this training process is evident -- as when you give it a large number of problems and their correct answers -- but sometimes the training is implicit.
When you “tag” a picture of your friend in Facebook, you are inadvertently giving Facebook a correct answer to a face recognition problem that helps train its software. Similarly, when you click on link in Google search, you identifying the most relevant answer and so Google is learning about your personal preferences so as to answer your next query better! In general, any action that you take on Facebook, a like, an emoticon, a phrase that you type as a post or a comment, is being used by the ANN to learn more about you so that it can predict your next behaviour -- which could be anything from accepting friend requests to clicking on ads!
Computers have been used for a long time to perform mathematical, process automation and other well structured tasks. But despite great increases in the speed and memory, tasks that are ambiguous or unstructured were never addressed adequately. Many people claimed that intelligence -- something evident in children or even animals -- is something that mere machines can never acquire. ANNs, that can learn, can be trained, has now helped us solve unstructured problems that were once considered intractable.
But unlike humans, machines can learn very fast -- working out 100,000 addition problems is almost impossible for a man but is trivial for a machine. So is the case with reading through all the articles in Wikipedia. If learning is the key to intelligence, as is commonly understood here, and if machines can learn much faster than humans, then does it mean that machines will become more intelligent than humans?
Tipping into the Singularity
In his 2000 bestseller, Malcolm Gladwell, defined the Tipping Point, as "the moment of critical mass, the threshold, the boiling point”, at which a small change, a sudden trigger, can suddenly usher in a huge change in the larger society. The phrase tipping point originates in the study of epidemics, and refers to the moment when a virus reaches a certain critical mass and then suddenly begins to spread at an accelerated rate. While Gladwell restricts himself to the analysis of social trends like crime and fashion, Ray Kurzweil, in a series of books on the technological singularity, predicts that human society as a whole is just about ready to transform itself in a way that is perhaps inconceivable today.
The trigger in this case is artificial intelligence or its alter ego, machine learning. The hypothesis is that society will pass through this technical singularity when non-biological intelligence transcends biological intelligence and changes the way not just how we live and locomote, but the way we think and what we believe. But is this hypothesis sustainable? Or is this another big hype -- that has been with us since the dawn of the science fiction era?
Ten years ago, Deep Blue, IBM’s chess playing computer had created history by beating the world champion, Gary Kasparov in 1996. Last year, AlphaGo, a computer based on an artificial neural network (ANN) , built by DeepMind, a Google company, that had “learnt” to play GO a game that is considered to be far more difficult for computers, had beaten Lee Sedol, one of the highest ranked professional players in the world. In the first week of January 2017, AlphaGo, playing anonymously under the handle of “Master” ran through China’s online GO playing websites and beat almost every top ranked player with contemptuous ease. Playing with inhuman speed, it eventually ended its unbroken 60-0 winning streak by beating Ke Jei, the reigning world champion. More than the defeat itself, its style, the strategies that it evolved on its own, are so different from that used by humans that the world champions were awestruck. After losing to Master, Ke, the current world champion admitted on social media : “After humanity spent thousands of years improving our tactics, computers tell us that humans are completely wrong, (and) I would go as far as to say that not a single human has touched the edge of truth of Go.” Playing in a manner completely different from humans, it bewildered opponents with apparently foolish moves that placed pieces at outrageously unconventional positions so that one player, Gu Li noted that “I can’t help ask, one day many years later, when you find your previous awareness, cognition and choices were all wrong, will you keep going along the wrong path or reject yourself.”
The key elements of the computer program were not written by humans but were discovered by the program itself as it learnt to play Go by observing millions online games, as explained earlier.
Very similar is the case of Interlingua, a new language that has evolved, or emerged out of Google’s automatic language translation software. After being trained to translate say, French to English and then English to Hindi, the software has learned how to translate from French to Hindi even though it was never “trained” to do so. This new language, or new way to represent ideas, has emerged not from the mind of programmers but from the neural network architecture that drives the software. This is an emergent phenomenon, similar to the appearance of language in a primordial human society.
Since its uncertain beginning in the 1960’s artificial intelligence technology took nearly 40 years to reach the level of maturity to beat a human chess champion, but the next big leap, when it learnt to play Go took only 10 years. This is the law of accelerating returns -- where key jumps in human or biological ability happen at increasingly shorter intervals, or where the gap between between innovations shorten exponentially as shown by Ray Kurzweil :
[ Graph #1 source Wikimedia Commons Ray Kurzweil]
This graph shows the time gap between key events in the history of human social evolution plotted against the time frame of history and it shows a nice linear slope because this is log-log graph where we have taken the logarithm of the X and Y values in the plot. Similar graphs drawn with key events identified by others show a very similar downward trend. [ for example this ]
However, if we convert the historical time frame to a linear scale, then drop in the time between two successive key events is clearly precipitous and this is what leads us to the concept of technological singularity.
Those who are familiar with Moore’s Law would know that the density of transistors on computer chips has been doubling every two years -- with a corresponding fall in prices and increase in computation power. This exponential pace of technology has been the driver behind the inexorable and astonishing growth of computers. While Moore’s law operates on computer chips and has been valid for the past 50 years, these charts show that human innovation has been accelerating exponentially across an even bigger time scale. Each human invention, like the ability to walk upright, usage of tools, language, agriculture, writing, printing, industrial machinery, electricity, computers, internet -- have actually reduced the time necessary to reach the next key invention. For example, years of calculations done by Kepler to track the movements of planets can now be done by school student on his desktop and with Google search, it is far easier, for someone to tap into the knowledge of others and create a new piece of technology than what it would have taken 100 years ago.
In another view of this same chart, we should see an exponential -- almost magical -- growth in the human ability to address problems that were thought impossible. After millions of years of being bound to the ground, man learnt to fly and then in 60 years he was on the moon and is today planning to go to Mars. So is the case with industrial machinery, telecommunications, healthcare and a host of allied fields. We are at, or very close to, an inflexion point -- a tipping point -- that will put us on a trajectory that leads to what seem to be a world of magic that may include immortality, omniscience or anything that does not violate the fundamental laws of physics!
A key characteristic of the tipping point would be the blurring of the borders between the biological and the non-biological world. Just as the industrial revolution merged the ability of man and machinery to create the superhuman capability to manage enormous quantities of materials and energy with, for example, cranes, excavators, rocket engines -- so does the digital revolution merge the ability of humans and computers to create a similar superhuman capability that manages equally enormous quantities of information. Continuing with this analogy we see that just as industrial machines today are far more muscular than humans in their ability to lift loads or travel faster and farther, so would the next generation of intelligent machines, powered by artificial super intelligence, would far outstrip the mental ability of humans -- as demonstrated by AlphaGo.
This then is the singularity, that point in time in human history that human society will soon pass through, after which the intelligence of machines will surpass that of humans. This superior intelligence, coupled with the ability to handle gigantic quantities of information at superhuman speed will lead to a cascading impact on the emergence of new ideas that “will abruptly trigger runaway technological growth, resulting in unfathomable changes to human civilization.”
This explosion of intelligence, this chain reaction where each level of intelligence gives rise to an even higher level of intelligence, exhibited either by machines or in a hybrid cyborg of man and machine, at faster and faster speeds or at shorter and shorter intervals was first postulated by John Von Neumann, a pioneer of, among things, the digital computer. Von Neumann had coined the term singularity but never lived to see technology reach anywhere near the level that matched his postulates. Subsequently, IJ Good, Vernor Vinge and eventually Ray Kurzweil have elaborated the concept. However equally well known people like Paul Allen, who co-founded Microsoft and Gordon Moore, whose Moore’s Law is a classic example of this exponential progress, have raised doubts about the plausibility of this concept. But with the rapid growth of artificial intelligence as evident in the spectacular success of AlphaGo, it seems that the singularity is indeed near and despite a wide range of predicted dates, the median value of its estimated arrival is the year 2040.
Once we pass through the singularity, how will the world look after 2040? Would this non-biological intelligence help create new, biological, non-biological or hybrid “life-forms”? Would these life-forms spread out across the solar system and the galaxy? But before we look at such big question, let us explore something that is of immediate concern -- the job market.
The Future of EmploymentWhen computers were first introduced in India, there was a lot of concern that this would lead to widespread unemployment. In West Bengal, and elsewhere, Communist labour “leaders” led huge agitations against this new technology with the slogan : automation must be stopped with rivers of blood, and employees of banks and commercial establishments went on strikes to prevent the installation of computers. But in reality the introduction of computers did not lead to any major disaster. The number of jobs lost to computers and automation was more than offset by the number of new jobs created not only in the IT industry, but also in many new age businesses that were based on computer technology. The number of people needed to use computers or write programs for computers was more than the number of people who became redundant. So while certain individuals were retired or laid off, the overall number of people who were gainfully employed increased and the economy responded with buoyant optimism.
That was then, what will it be now? Will the new technology based on powerful instances of artificial intelligence create more jobs than it destroys? Can our earlier experience with an earlier generation of technology be extrapolated in time, through the technological singularity where machines become more intelligent than men? Unfortunately, the answer to both questions seems to be “No”.
While robots are not yet in widespread use in India, we can look at what is happening in other countries that are ahead of us in this path. Between 2000 and 2010, 5.6 million jobs were lost in the US and Canada. Of this only 15% were lost to overseas competitors -- mainly China, while the other 85% were due to, what is euphemistically referred to as, “productivity growth”. This means that humans were replaced by machines or robots. However this decline in employment did not result in lower production. On the contrary, in the last 20 years, the value added by US factories has, after adjusting for inflation, grown by nearly 40% to reach a record US$ 2.4 trillion. On the street, McDonalds’ introduction of self-service kiosks in response to popular pressure for $15/hour wage is a vivid example.This is job-less growth, where the economy expands but without a corresponding increase in employment and is perhaps one reason why lots of middle-class Americans who have been rendered unemployed and unemployable in the rust belt states have aggressively turned against the establishment and voted for Donald Trump.
Many Trump supporters believe that foreign countries are stealing their jobs but the situation is not greatly different there either. The BBC has reported that Foxconn, the Chinese company that manufactures products on contract from Apple and Samsung, has recently replaced 60,000 workers with robots and many other companies in the Kunshan region, where Foxconn factories are located, are likely to follow suit. Since 2013, companies in China have purchased more industrial robots than in any other country, as thousands of companies are turning to automation in a robot-driven automation drive that has been backed by the Chinese government in a desperate bid to remain competitive in manufacturing. Can the “Make-in-India” movement avoid this? Unlikely, because no army can stop an idea whose time has come!
The services industry, that is of more immediate concern to India is no better off. Thanks to better voice and natural language recognition techniques, call center operators -- the backbone of India’s BPO / ITES success story -- can be replaced by artificial intelligence driven bots that can do a far better job in patiently listening to customer problems and offering solutions. In fact, a whole range of service jobs are now at risk and these include but are not limited to cooks and chefs, medical doctors, surgeons, pharmacists and pathology laboratory staff, security guards, retail salespersons, receptionists, bar-tenders, farmers, truck, bus and taxi drivers and even those who perform unstructured tasks like journalists, accountants and insurance claims adjusters. This may sound too futuristic and science-fictionesque but as we have seen earlier, the fierce of acceleration of change makes it impossible to wish away these dire, Cassandra-like predictions on the future of employment. Amazon Go, a new kind of store that has no human employees is an example of what a typical retail store could be like in the future. It’s Just-Walk-Out technology, allows a customer to walk in, pick up products from the shelves, look at it, return it if necessary, then simply walk out without any formal check out procedure and yet be billed for it automatically on his credit card. This is not magic, this is not the future. This is here and now. Such a store is actually operating in Seattle on a trial basis for Amazon employees and it is simply a matter of time before it becomes a mainstream technology, just like Google’s technology for driverless cars is proliferating across the US and Europe and is about to be tested even by Tata Elxsi in Bengaluru.
When bank workers were replaced by banking software there was a big need for computer programmers to build and maintain banking software. When retail stores were replaced by e-commerce sites they needed people to build and maintain e-commerce software and also an army of delivery boys to cater to an expanding clientele in the previously unreachable mofussil areas. But when jobs are lost to robots and artificial intelligence systems, the number of jobs created is only a fraction of those lost. Unlike computer software, industrial robots are not built by humans but by other robots that are designed and programmed by a couple of very smart humans. So a large number of low end jobs are replaced by a few, high end, specialist jobs. This makes eminent economic sense for both the users and manufacturers of robots but leaves the newly unemployed worker distressed and angry.
While robots can always build robots, it was thought that at least the programming of such automated systems would be done by a few highly skilled humans -- but even such relief may be short lived. In January 2017, the MIT Technology Review has reported that Google is building systems that not only demonstrate AI in, say driving cars, but -- like a snake, recursively, swallowing its tail -- actually builds the the next-generation systems that demonstrate more AI. This is the eureka moment, comparable to the event of the DNA molecule being able to replicate itself and thus define the emergence of physical life. As AI systems build more advanced AI systems, they can take on a life, a non-biological life, of their own. So the need for human beings becomes even more insignificant and inconsequential in the economic systems of tomorrow’s world.
But then what do we do with humans who are no more necessary for the economy to generate goods and services? for whom would the “economy” generate goods and services? and who will pay for these goods and services? These are difficult questions that will crop up as humanity moves into an uncharted territory of the post-singularity era. An obvious way to contain the rising tide of resentment against unemployment would be to have strong social security systems like Uniform Basic Income (UBI) or even NREGA that will push money into people’s pockets without them having to do any work -- because there really is no useful work left that they could do.
Does this mean that people will simply stay at home and play cards? Or will they get drunk and create mischief? An idle mind could be the devil’s workshop. What kind of sociological and psychological problems will this lead us towards? Once again we have questions but hardly have any definitive answers. In the utopian scenario, we envisage that man will increasingly be involved in cerebral, cultural and entertainment activities -- art, music, literature, physical sports, cinema, virtual and augmented reality games, sex -- while machines, robots and AI systems will do the “dirty” job of keeping the economy running so that it generates a distributable surplus. The other, dystopian scenario would be a descent into anarchy as the gap between the unemployable poor and the talented rich becomes bigger and more bitter. We are looking at gated communities, or huge walled towns, well endowed with sustainable sources of food and energy, managed efficiently with high technology and defended with highly effective robotic systems against an angry, anarchic and violent outer world ruled by rogues and brigands.
These two extremes that we can atavistically think of are reminiscent of the middle ages -- the first, benevolent scenario resembles a society of serfs and noblemen while the second, malevolent scenario reminds us of medieval cities being islands of civilisation and governance in an otherwise anarchic and lawless countryside. But perhaps with the emergence and eventual predominance of artificial intelligence and non-biological “life” forms, our existing models of social behaviour will cease to be reliable predictors of hybrid, human-machine civilisation that we cannot envisage at the moment but are about to bequeath to ourselves anyway.
The accelerating pace of technology cries “Havok” and has let slip the dogs of rapid and irreversible change. The genie is out of the bottle and new “life” forms, with a different type of intelligence, is about to be released first into the economy, then into society and finally into the planet as a whole. Would this wipe out humans? Or would man evolve and adapt itself to co-exist with this new “species”? Would he retain his position as the master at the top of the global ecosystem? Or would this position be taken over by a machine or a hybrid cyborg that combines the best -- or worst -- of both man and machines? The answer lies in the womb of futurity.
This article originally appeared in Swarajya, the magazine that reads India right