THURSDAY, 23 MAY 2019
IN MAY 1997, IBM’s computer program ‘Deep Blue’ infamously defeated chess world champion Garry Kasparov in a set of six highly anticipated games. In a curious case of repeated history, DeepMind’s program ‘AlphaGo’ did the same for the ancient Chinese board game of ‘Go’ 19 years later, beating 18-time world champion Lee Sedol 4–1 in March 2016.While beating another board game’s top player after two decades hardly seems like significant progress, the underlying differences between ‘Deep Blue’ and ‘AlphaGo’ reveal a profound paradigm shift in recent computing. AlphaGo is proof that neuroscience has become as vital as electronics and mathematics in computer science for solving many modern problems.
AlphaGo is an example of a ‘neural network’ program. In the brain, neurons are cells which take in, process, and transmit electrical signals. The chemical processes inside these neurons and the links between them define how humans think, feel and move. This led to a bold new idea - what if computers could model these neurons?
In 1943, neuroscientist Warren McCulloch and mathematician Walter Pitts proposed a way to simulate the brain’s neurons. They represented the brain’s chemical processes as mathematical functions, and electrical signals between neurons as on-off signals that were passed from function to function. By manually writing these functions and linking them together, McCulloch and Pitts created the first artificial neural networks.
However, something was missing from their model—learning. What makes us most human—the ability to learn—stems from the malleability of neurons. Their model did not consider that neural connections can strengthen, diminish or disappear entirely as the brain accumulates experience, and learns what is right and wrong. In 1957, Rosenblatt was the first to fill this gap by allowing the artificial neural connections to adapt over time. By enabling more varied calculations inside neurons, connections between neurons also became more mathematically complex and far more powerful. With these changes, Rosenblatt successfully simulated a neural network that could identify a triangle in a 20x20 image. Rosenblatt had remarkable ambitions for his newly-dubbed ‘perceptrons’. He declared that his perceptrons could be “fired to the planets as mechanical space explorers” in the near future. While this certainly captured the public’s imagination, he soon faced the reality that computers of his time simply weren’t powerful enough and his model was not thorough enough to properly model the human brain.
The relationship between neuroscience and computer science began to stagnate after Rosenblatt’s discoveries. More focus was placed on sheer computing power, and with good reason—the aspirations of neuroscientists simply could not be met by computers of the day. Mathematics took the lead in advancing research into neural networks: refining the layout of neurons; optimising computations within the neurons; and developing new techniques to imitate the brain’s ability to learn. For many decades, there ironically wasn’t room for neuroscience in neural networks.
However, these researchers’ efforts are now paying off in the modern era. Computing power has increased more than a billion-fold since Rosenblatt’s era, and new chips, dubbed ‘Tensor Processing Units’ (TPUs) are now being specifically designed to improve neural networks. These advancements have given this research of the past new life and caused a surge in neural network performance.
In the ‘ImageNet Large Scale Visual Recognition Competition’ (ILSVRC), a competition testing neural networks’ ability to identify animals and objects in a set of thousands of images, the accuracy of the winning algorithm has increased from roughly 71.8% in 2010 to a staggering 99.98% in 2017.
Many of the problems that we face in computer science today are fundamentally human problems. Image recognition, speech recognition and driving, for example, are all very human tasks, and we are aiming to create a system which can match, or even exceed, our own abilities. It should therefore be of no surprise that neural networks are already at the forefront for what is possible, as they are most closely linked to how we approach these problems ourselves.
Nevertheless, there are still many problems that neural networks cannot solve. For instance, today’s image recognition neural networks are plagued by issues with ‘adversarial tests’, which use images specifically designed to confuse the model.
Rosenfeld, Zemel and Tsotsos showed this in an August 2018 study by inserting an elephant drawing into an image of a man in his house. Depending on the position of the elephant, the neural network may have ignored the elephant, misinterpreted the elephant as a chair, or even have become confused about other objects in the image which were correctly identified before the elephant was introduced.
Tsotsos, the neuroscientist of the group, explains how this demonstrates that large sections of brain function are still under-utilised in computer science. A human, upon seeing such an image, would first recognise that the elephant is out of place and proceed to examine the elephant and the rest of the image separately. In essence, humans have learnt to do a double take in order to understand images more accurately and more efficiently. Even today’s best neural networks haven’t yet made such logical leaps.
Back in the world of board games one question still remains: What sets DeepMind’s neural networks apart from Deep Blue? Deep Blue was the product of hundreds of hours encoding moves, positions and strategies. A team of engineers, computer scientists and chess grandmasters at IBM harnessed computers’ sheer power to calculate the best possible move in any turn. But even computers have their limits, and it was clear that this strategy would not work again for the complexities of Go.
DeepMind’s newest offering, ‘AlphaZero’, is much simpler. It is given the rules of Go, and nothing else. It is left to its own (literal) devices, it repeatedly plays itself over and over again, and it learns from every win and loss. Within 21 days, AlphaZero becomes quantifiably superhuman, beating Go world champions consistently.
What’s most remarkable about this system, though, is that AlphaZero can be given any ruleset. Given the rules of chess and just nine hours to train itself, AlphaZero is able to beat not just humans, but also the world-leading programs that have dominated the chess world since 1997.
The fatal flaw of Deep Blue is that it needs to be taught by someone, and the better the teacher, the better the result. In contrast, AlphaZero is not taught. It is not instructed what to do, and is thus not limited by the abilities of its teachers. Instead of caring about what humans think would work best, it learns everything it needs all by itself.
In many ways, you can already see humanity in artificial intelligence systems like AlphaZero. Just like humans, AlphaZero is not designed to solve a single problem. It is designed to absorb information, and to learn from its mistakes just as we learn from our own. And why does AlphaZero seem so human sometimes? It has been made entirely in our image. The processes of trial and error, adaptation and learning are inherent to the neural networks that drive both our brain and AlphaZero. It will be systems like these which will be able to pick up fundamentally human skills like speaking language fluently, recognising objects at a glance, and understanding nuanced facial expressions.
It is not clear just how far neuroscience and computer science will take us. Perhaps we will never understand the brain well enough to make a computer completely in its likeness. In any case, neuroscience will continue to inform and inspire the development of neural networks, and will no doubt have a profound impact on the evolution of computer science.
Charles Jameson is a first year Computer Scientist at Queen's College.
Banner artwork by Nah Yeon Lee.
