Is the Brain a Good Model for Machine Intelligence?

COMMENT

equivalent to Turing’s finite-state machine with an infinite tape and a finite symbol set, and that does computation. In 1943, Warren McCulloch and Walter Pitts2 noted the “all-or-none” nature of the firing of neurons in a nervous system, and suggested that networks of neurons could be modelled as logical propositions. They POTTS ANDY BY ILLUSTRATION modelled a network of neurons as circuits of logic gates, noting that these may “compute only such numbers as can a Turing machine”. But more, they proposed that everything at a psychological level happens in these networks. Over the decades, such ideas begat more studies in neural networks, which in turn begat computational neuroscience. Now those metaphors and models pervade explanations of how the brain ‘computes’. But these binary abstractions do not capture all the complexities inherent in the brain. So now I see circles before my eyes. The brain has become a digital computer; yet we are still trying to make our machines intelligent. Should those machines be modelled on the brain, given that our models of the brain are performed on such machines? That will probably not be enough. When you are stuck, you are stuck. We will get out of this cul-de-sac, but it will take some brave and bright souls to break out of our circular confusions of models. DEMIS HASSABIS Is the brain a good model Model the brain’s algorithms for machine intelligence? Neuroscientist, computer-game producer and chess master, To celebrate the centenary of the year of Alan Turing’s University College London

birth, four scientists and entrepreneurs assess the Alan Turing looked to the human brain as divide between neuroscience and computing. the prototype for intelligence. If he were alive today, he would surely be working at the inter- section of natural and artificial intelligence. model. He abstracted the actions of a human Yet to date, artificial intelligence (AI) RODNEY BROOKS ‘computer’ using paper and pencil to per- researchers have mostly ignored the brain form a calculation (as the word meant then) as a source of algorithmic ideas. Although Avoid the cerebral into a formalized machine, manipulating in Turing’s time we lacked the means to look symbols on an infinite paper tape. inside this biological ‘black box’, we now blind alley But there is a worry that his version of have a host of tools, from functional mag- computation, based on functions of inte- netic resonance imaging to optogenetics, Emeritus professor of robotics, gers, is limited. Biological systems clearly with which to do so. Massachusetts Institute of Technology differ. They must respond to varied stimuli Neuroscience has two key contributions over long periods of time; those responses to make towards progress in AI. First, the I believe that we are in an intellectual cul-de- in turn alter their environment and subse- many structures being discovered in the sac, in which we model brains and computers quent stimuli. The individual behaviours of brain — such as grid cells used for naviga- on each other, and so prevent ourselves from social insects, for example, are affected by tion, or hierarchical cell layers for vision having deep insights that would come with the structure of the home they build and the processing — may inspire new computer new models. behaviour of their siblings within it. The first step in this back and forth was Nevertheless, for 70 years, those people 1 TURING AT 00 made by Alan Turing. In his 1936 paper working in what is now called computa- A legacy that spans science: laying the foundations of computation, tional neuroscience have assumed that the nature.com/turing Turing used a person as the basis for his brain is a computer — a machine that is

462 | NATURE | VOL 482 | 23 FEBRUARY 2012 © 2012 Macmillan Publishers Limited. All rights reserved COMMENT algorithms and architectures. Second, anatomy or physiology. Even biologically Two of the many fundamental differences neuroscience findings may validate the plau- inspired approaches such as cellular autom- between the brain and the computer are sibility of existing algorithms being integral ata, genetic algorithms and neural networks memory and processing speed. The analogue parts of a general AI system. have only a tenuous link to living tissue. of long-term memory in a computer is To advance AI, we need to better under- In 1944, Turing confessed his dream of the hard disk, which can store practically stand the brain’s workings at the algorithmic building a brain, and many people continue unlimited amounts of data. Short-term infor- level — the representations and processes in that endeavour to this day. Yet any neuro mation is held in its random access memory that the brain uses to portray the world biologist will view such attempts as naive. (RAM), the capacity of which is astronomical around us. For example, if we knew how How can you represent a neuronal synapse — compared with the human brain. Such quan- conceptual knowledge was formed from per- a complex structure containing hundreds of titative differences become qualitative when ceptual inputs, it would crucially allow for the different proteins, each a chemical prodigy in considering strategies meaning of symbols in an artificial language its own right and arranged in a mare’s nest of “Signals in for intelligence. system to be grounded in sensory ‘reality’. interactions — with a single line of code? We the brain are Intelligence is mani- AI researchers should not only immerse still do not know the detailed circuitry of any transmitted fested by the ability to themselves in the latest brain research, but region of the brain well enough to reproduce at a snail’s learn. Machine-learning also conduct neuroscience experiments to its structure. Brains are special. They steer us pace.” practitioners use ‘stat address key questions such as: “How is con- through the world, tell us what to do or say, istical learning’ which ceptual knowledge acquired?” Conversely, and perform myriad vital functions. Brains requires a very large collection of examples from a neuroscience perspective, attempt- are the source of our emotions, motivation, on which to generalize. This ‘frequentist’ ing to distil intelligence into an algorithmic creativity and consciousness. Because no one approach to probabilistic reasoning needs construct may prove to be the best path to knows how to reproduce any of these features vast memory capacity and algorithms that are understanding some of the enduring mys- in an artificial machine, we must consider at odds with available data on how the brain teries of our minds, such as consciousness that something important is missing from works. For example, IBM computer Watson and dreams. the canonical microchip. needed to consume terabytes of reference Brains differ from computers in a number material to beat human contestants on Jeop- of key respects. They operate in cycles rather ardy!. Volvo’s pedestrian-detection system DENNIS BRAY than in linear chains of causality, sending (developed by Mobileye) learned to identify and receiving signals back and forth. Unlike people by using millions of pictures. In both Brain emulation the hardware and software of a machine, the cases, the human brain is considerably more mind and brain are not distinct entities. And parsimonious in the reliance on data — some- requires cells then there is the question of chemistry. thing that does not constrain the computer. Living cells process incoming sensory In terms of processing power, the brain Department of Physiology, information and generate not just electri- can reach about 10–50 petaflops — equiva- Development and Neuroscience, cal signals but subtle biochemical changes. lent to hundreds of thousands of the most University of Cambridge Cells are soft, malleable and built from an advanced Intel Core i7 CPUs. Yet signals essentially infinite variety of macromolecular in the brain are transmitted at a snail’s pace Machines can match us in many tasks, but species quite unlike silicon chips. Organisms — five or six orders of magnitude slower they work differently from networks of nerve encode past experiences in distinct cellular than modern CPUs. This huge difference in cells. If our aim is to build machines that are states — in humans these are the substrate communication speed drives vastly different ever more intelligent and dexterous, then we of goal-oriented movements and the sense architectures. should use circuits of copper and silicon. But of self. Perhaps machines built from cell-like The brain compensates for the slow if our aim is to reproduce the human brain, components would be more like us. signal speed by adopting a hierarchical paral- with its quirky brilliance, capacity for multi- lel structure, involving successive layers with tasking and sense of self, we have to look for increasing receptive field and complexity. By other materials and different designs. AMNON SHASHUA comparison, a computer architecture is usu- Computers outperform us in complex ally flat and, because of its much faster clock mathematical calculations and are better at Speed will trump rate, can employ brute-force techniques. storing and retrieving data. We accept that Computer chess systems such as Deep Blue they can beat us at chess — once regarded brain’s advantages use pattern-recognition strategies, such as as the apogee of human intellect. But the libraries of opening moves and completely success of a computer called Watson in US Sachs Professor of Computer Science, solved end-games, complemented by their television quiz show Jeopardy! in 2011 was Hebrew University of Jerusalem, and ability to evaluate the outcomes of some a nail in the coffin of human superiority. co-founder and chairman of Mobileye 200 million moves per second. This is way The machine beat two human contestants beyond the best grandmaster. by answering questions posed in colloquial The saying that “people who are really An intimate understanding of how cogni- English, making sense of cultural allusions, serious about software should make their tive tasks are performed at an algorithmic metaphors, puns and jokes. If Alan Turing own hardware”, attributed to computer level would allow artificial intelligence to had been given a transcript of the show, scientist Alan Kay in the 1980s, still rings grow in leaps and bounds. But we must bear would he have spotted the odd one out? true today. The idea that the function and in mind that the vastly different architec- Watson may be the latest vindication of form of computing architecture should serve ture of the computer favours strategies that Turing’s view of intellectual processes as a each other is at the root of algorithms in make optimal use of its practically unlimited series of logical states. But its internal work- signal processing, image rendering, gaming, memory capacity and brute-force search. ■ ings are not based on the human brain. Broad video compression and streaming. I believe 1. Turing, A. M. Proc. Lond. Math. Soc. s2-42, similarities in organization might be imposed that it is also true for the human brain — 230–265 (1936–37). by the nature of the task, but most software meaning that the brain does not implement 2. McCulloch, W. S. & Pitts, W. H. Bull. Math. Biophys. engineers neither know nor care about ‘intelligence’ in the same way as a computer. 5, 115–133 (1943).