MACHINES WITH A MIND OF THEIR OWN

MACHINES WITH A MIND OF THEIR OWN

A ‘thinking’ machine isn’t just the stuff of sci-fi movies. With computer science and neuroscience working together to simulate the human brain, a breakthrough in artificial intelligence may not be too far away

    Last October, members of an audience at a conference in Tianjin, China, gasped when they heard Microsoft chief research officer Rick Rashid address them in Mandarin. He would speak in English, pause, and the Chinese translation would come on, in his own voice. A machine was converting spoken English into text English, translating that to text Chinese and then converting it to spoken Chinese using Rashid’s own voice characteristics.
    Chinese is a difficult language, and it is also structured differently from English. So the task was very difficult. The tech world hailed the new software as a breakthrough. After a long time it looked as if progress was being made towards liberating Artificial Intelligence (AI) — that is, machines that behave like humans — from the sci-fi movie cage.
    There had been milestones like that before. One of the most publicized of such events was in 1997 when an IBM machine called Deep Blue beat world chess champion Garry Kasparov two wins to one, with three draws. The refrigerator-sized computer could analyze 200 million positions in a second. It had a database of 700,000 grandmaster games to draw upon.
    The world hailed this event as a historic moment — the first real sign that a machine could be more ‘creative’ than a human. But, ultimately, what was Deep Blue doing? It was utilizing brute force power to ram through millions of calculations to beat an undoubtedly creative and trained mind like Kasparov’s. In 2006, a German software called Deep Fritz contained in just two Intel processors beat the new world champion Victor Kramnik. Again brute force, but contained in a much more elegant machine.
    More recently, there was Watson, a machine built by IBM specially for playing the American TV game show Jeopardy! in which contestants have to give the right questions to answers read out to them. In 2011 Watson took on two of the game’s alltime biggest winners Brad Rutter and Ken Jennings – and beat them soundly. It contained 200 million pages of content, including the full text of Wikipedia. It had four terabytes of disk storage and 16 terabytes of RAM (memory) storage. It could process 500 gigabytes of data, equivalent to a million books in one second. (Tera is trillion, giga is billion.)
    So was it brute force power again? It was, but there was more to it. Watson had broken the language barrier between humans and computer machines. Humans speak in natural language rightly assuming that the listener will fill in. Scientists had struggled against this barrier for decades because it seemed impossible to build a machine with a databank that could cover all the quirks and styles of speech. But Watson had broken through in what the techies call ‘Natural Language Processing’, although it did falter once or twice in the contest.
    Machines have undoubtedly come a long way towards human intelligence in the past half a century or so. Besides high-visibility achievements like Deep Blue and Watson, computing machines have developed enormous, unimaginable speeds that no human could dream of. Simultaneously, size and energy consumption have gone down drastically. Many strides have been made in bridging the machine-human chasm — from voice and face recognition to bionic artificial limbs that respond to nerve impulses.
    But there is still a long way to go. When can a machine be called intelligent? There is much controversy on this, and rather bizarrely, every new advance appears in retrospect to not yet reach the level of ‘intelligence’. It is generally agreed that the gold standard for declaring a machine intelligent is that it should pass the Turing test.
    Alan Turing, British code-breaker and father of computational theory, in a paper in 1950 proposed that if a human being cannot distinguish between a machine and a human through interaction, then that machine would be intelligent. This is how this thought experiment is visualized: in one room there is a computer and a human with a computer; in another sits the interrogator, a human with a computer connected to both computers in the other room. The interrogator interacts with both computers and tries to guess which is run by a human and which one is running by itself. When the computer manages to fool the interrogator into thinking it is human, that day the intelligent machine would be born.
    Despite all the spectacular advances, no machine has come anywhere near
passing the Turing test.
    Here is a recent innovation that brings out both the achievements and the limits of intelligent machines. Stanford University Electrical Engineering professor Andrew Y Ng and Google fellow Jeff Dean, working at Google X, a research facility at an unknown location in the Bay area, San Francisco, reported last year that they had developed a 16,000 array of processors. The system was fed 10 million YouTube thumbnails so that it could ‘watch’ and categorize what it saw into 22,000 categories like ‘cats’, ‘humans’, ‘cars’, and so on.
    The researchers called their system an “unsupervised neural network”, that is, it was modeled on the way neurons (brain cells) are organized in the brain, hierarchically and connected to each other. “Unsupervised” because it was not told specifically to "identify cats” or anything like that. It was just built to analyse and categorise. Because of the sheer number of cute cat videos on the Internet, the system started identifying cats and slotting them. Similarly it identified ‘humans’. No computer system ever had been able to do this before in “unsupervised” conditions. It looked as if a “learning machine” was finally coming through.
    But here’s the catch: the system had a success rate of 16%. This is 70% more than what had been achieved ever before. Still, it is way short of what a human eye (with the visual cortex of the brain backing it up) can do. The human visual cortex, located at the very back of the brain, has a million times more connections than the one achieved in the Google system, the researchers themselves admitted.
    “It’d be fantastic if it turns out that all we need to do is take current algorithms and run them bigger, but my gut feeling is that we still don’t quite have the right algorithm yet,” Andrew Ng told The New York Times.
    This approach of modelling computers on the human brain is gaining increasing popularity and some modicum of success. The seat of thinking, language, motor functions, spatial and sensory perception in the brain lies in the neocortex, the uppermost layer of the brain consisting of grey cells. It is found only in mammals, and the biggest neocortex in relation to the rest of the brain is found in humans. The neocortex is made of six layers and it is thought that this is a hierarchical system — neurons start reacting at the lowest (most inside) layer and each succeeding layer refines the activity to produce the final ‘thought’.
    One of the leading figures in the field of building neural network based machines is Geoffery Hinton, a professor at the University of Toronto, who spends half his time at Google. He has developed simple models of virtual neurons connected to each other and layered like the human neocortex.
    Of course, neuroscience itself is, still “a bit like physics before Newton,” as Bruno Olshausen, director of the Redwood Center for Theoretical Neuroscience at the University of California-Berkeley put it in the journal Scientific American. Scientists think that they have understood only about 15% of how the visual cortex works — and that is just one of the functions of the brain.
    But, as the functioning of the brain is revealed, spurred by advances in imaging technology, the feedback loop with computer scientists will expand rapidly. And, so will the prospect of thinking machines.