Turing award winner Prof. Leslie Valiant believes that biological evolution and the way computers learn are governed by the same set of rules!
For decades, scientists have tried to make robots more like human beings by designing computer algorithms that allow them to learn and become smarter.
In fact, similar kinds of biological algorithms might exist in people and govern not only how we learn and act but also how our species evolved.
That is the firm belief of Professor Leslie Valiant, whose ground-breaking research has been fundamental to the development of machine learning, artificial intelligence and the broader field of computer science.
“You think of an algorithm as something running on your computer, but it could just as easily run on a biological organism,” said the 67-year-old in an interview in January (2016) with Quanta Magazine, which reports on developments in mathematics and the physical and life sciences.
“If one has a more high-level computational explanation of how the brain works, one would get closer to having an explanation of human behaviour that matches our mechanistic understanding of other physical systems.”
Teaching machines to learn
One of the speakers at the recently concluded Global Young Scientists Summit@one-north 2017 (GYSS 2017), Prof Valiant has long blurred the lines between the computer and life sciences in his work.
In 1984, inspired by the fact that human beings appear to be able to learn new concepts without needing to be programmed explicitly, he published a seminal paper outlining the conditions under which a machine could also be said to “learn”.
This paper, titled “A Theory of the Learnable”, provided both a general framework for machine learning as well as concrete computational models.
His approach, called the “Probably Approximately Correct” (PAC) model, posits that a machine can create useful generalisations by examining an array of examples.
For instance, it might be able to determine the characteristics of human beings by analysing examples of mammals that are labelled either “a human being” or “not a human being”.
From these examples, it might develop the classification that human beings are “warm-blooded, have opposable thumbs and give birth to their young”. It would then be able to use this algorithm to decide if animals it sees in the future are human beings or not.
Of course, this algorithm is not entirely correct. Chimpanzees, too, have those characteristics, and if the samples did not include chimpanzees, the machine might go on to mistake chimpanzees as human beings.
There is also the minute chance that the examples labelled “a human being” all have fair hair, thus causing the machine to add the incorrect generalisation that human beings must have fair hair.
While the machine’s algorithms are unlikely to be always and entirely correct, with enough examples, they are likely to be probably and approximately correct – hence the name of the model.
The PAC model has become one of the most important contributions to machine learning, and is the foundation of the modern field of computational learning theory, where scientists study the design and analysis of machine learning algorithms.
For this and other trail-blazing work, Prof Valiant was awarded the 2010 Turing Award, which is regarded as the Nobel Prize in computing.
The international Association for Computing Machinery (ACM), which bestows the award, said in its citation that Prof Valiant “brought together machine learning and computational complexity, leading to advances in artificial intelligence as well as computing practices such as natural language processing, handwriting recognition and computer vision”.
In fact, “mainstream research in artificial intelligence has embraced his viewpoint as a critical tool in designing intelligent systems,” the ACM added.
The Algorithms to Life
The PAC model’s usefulness might also not be limited to machines. More recently, Prof Valiant, who is the T Jefferson Coolidge Professor of Computer Science and Applied Mathematics at Harvard University in the United States, has expanded his theory to include biological evolution.
He believes, for instance, that Darwin’s theory of evolution is convincing but incomplete. It does not explain, among other things, the rate at which evolution occurs.
“Amazingly, although the evidence that evolution has taken place is overwhelming, we still do not have a quantitative explanation of how it could have proceeded as fast as it has on Earth. Understanding the speed of evolution is a fundamental goal,” he said.
“The PAC framework spells out what biological details need to be understood before such an explanation might be derived, and offers methods of analysis related to the speed once those details are known.”
Prof Valiant believes that if biologists and computer scientists collaborated more, they might be able to unlock the algorithms used in biology that would explain not just evolution but human behaviour as well.
In his 2013 book, also titled “Probably Approximately Correct”, he notes that if human beings are shaped entirely by evolution before conception and by learning afterwards, all of our characteristics, whether biological or psychological, will have been determined by adaptive mechanisms.
He came up with the concept of an “ecorithm”, which is essentially a learning algorithm that runs on any system that can interact with its physical environment.
“There is a clear argument for the statement that all aspects of an individual’s behaviour are controlled by the joint influence of the evolution of the species and what the individual has learned. If evolution and learning are both forms of ecorithms, then the study of ecorithms offers a unified approach to understanding both,” he said.
“For example, the many ways in which human reasoning is prone to mistakes, as studied by psychologists over the decades, could be reflections of the way humans learn and represent information in the brain.”
Cracking the codes of such ecorithms could also help scientists to develop more advanced robots that can better learn from their environment, allowing the machines to evolve in a manner similar to people and become more useful.
While popular culture is replete with depictions of super-intelligent machines run amok, Prof Valiant does not believe in such doomsday scenarios.
“I regard intelligence as made up of tangible, mechanical and ultimately understandable processes,” he said in the Quanta Magazine interview.
“We will understand the intelligence that we put into machines in the same way that we understand the physics of explosives – that is, well enough to render their behaviour predictable enough that in general they don’t cause unintended damage.
“I’m not so concerned that artificial intelligence is different in kind from other existing powerful technologies. It has a scientific basis like the others.”
The above article was written by Feng Zengkun for National Research Foundation, the organizer of GYSS@one-north.