Next January 12th marks the 25th anniversary of the literary debut of HAL,
computer star of 2001:A Space Odyssey and enigmatic serial killer par
excellence. A sensitive soul with little stomach for telling lies, HAL ends up
killing all but one of his crewmates for fear of being unplugged.
A more vivid warning of the perils of giving computers emotions it’s hard to
imagine. And yet, when Rosalind Picard celebrates HAL’s birthday next year in
the pages of a commemorative book, it won’t be to dwell on the dark side of
Arthur C. Clarke’s tale. Instead, this professor of computer science and
communication at the Massachusetts Institute of Technology will tip her glass to
a rosier, Earth-bound vision of what the world would be like if computers were
truly emotional.
And what a world she envisions. Keyboards, mouse pads, swivel chairs, you
name it, says Picard, they might all be redesigned one day to soak up those
tell-tale clues about our emotional state—from typing rhythm and finger
pressure to skin temperature, pulse, posture and facial expression. Armed with
all that information, why not develop a new generation of “affective
ł¦´Çłľ±čłÜłŮ±đ°ů˛ő”?
Advertisement
Think of the benefits. Wearable devices could relay timely clues about your
emotional state to your partner or spouse. Emotionally sensitive cars could
“know” when you’re stressed to breaking point and automatically switch on
relaxing music. Computerised tutors would behave sensitively to the needs of
their pupils. And simulated emotional states in computer models could help
would-be therapists learn more about possible ways of dealing with disturbed
clients, just as flight simulators help to train pilots.
Flights of fancy they may be right now, but behind the breathless vision lies
a serious shift in scientific thinking about emotions and the role they could
play inside the machines of the future. Much of the impetus comes from a growing
recognition that Mr Spock got it wrong: far from acting in opposition to reason,
human emotions are essential to decision making, planning and judgment. Any
lingering loyalties to the Spock view are being swept away by behavioural
studies of people who lack the ability to feel the full range of human emotions.
The question for computer scientists thus seems obvious: will machines without
emotions and feelings ever be able to think and plan effectively?
Based on behavioural research findings, Antonio Damasio, a neuroscientist at
the University of Iowa, believes not. In his book, Descartes’ Error, Damasio
recounts the story of “Elliott”, a perfectly normal man who in his thirties
undergoes surgery for a brain tumour. The operation is successful, but
afterwards Elliott’s life unravels in a series of personal and economic
disasters. Marriages, jobs, relationships, business ventures—all fail.
Intellectually, Elliott is unimpaired. IQ and memory tests reveal nothing
abnormal. Only after much probing with ingenious psychological and physiological
tests do Damasio and his colleagues discover the root of Elliott’s
problem—an inability to make effective decisions or plan ahead, even for a
few hours.
Eventually the researchers trace this myopic indecisiveness to a curious
absence of feeling, itself the result of damage to the frontal part of the
brain’s cortex. Elliott “knows” but cannot “feel”. When confronted with pictures
of people injured in gory accidents, he knows intellectually that he should feel
distressed—but he doesn’t actually feel distressed: his body and brain
don’t undergo the usual neural and chemical changes that accompany such
emotions.
Without these emotional changes to guide his thought processes, concludes
Damasio, life for Elliott is a hell of indecision. Yes, he can mull over every
option ad infinitum; but when it comes to experiencing the subtle internal
values and biases of feeling necessary for actually choosing between the
options, “gut feelings” or “instincts” are just plain missing. Elliott, in
Damasio’s own words, is “irrational concerning the larger framework of
˛ú±đłó˛ą±ąľ±´ÇłÜ°ů”.
Other, even more subtle neurological problems have recently come to light
indicating a vital role for feelings in our cognitive lives. Take the curious
impact of emotions on memory and learning. Most of us are better at recalling
stories and events charged with trauma or disturbing detail. But last year,
Larry Cahill, a neuroscientist at the University of California at Irvine, and
his colleagues studied a patient whose powers of recall turned out to be
completely unaffected by the emotional content of the story. The patient
suffered from a rare hereditary disease that had damaged neural circuitry vital
for evaluating the emotional significance of our experiences. In so doing, the
disease robbed the patient’s memory systems of any ability to discriminate
between the traumatic and the benign.
The broader message from such cases is that if too much emotion makes people
irrational and undiscriminating, then so too does too little. And it is this
insight that computer scientists are now taking to heart. Picard, for one, is
convinced that if computers are to be truly effective at decision-making, “they
will have to have emotion-like mechanisms working in concert with their
rule-based systems”. Pure reason, she believes, is fine as a Platonic ideal,
“but in successful cognitive systems, it is a logical howler”.
The big question, of course, is how to put such notions into practice. How do
you give computers the kinds of “feelings” that would make them less like
Elliott and more like normal people?
Picard believes that computers must first be taught to recognise emotional
expressions. “Humans almost always respond to highly emotional events with
emotion, either to empathise or counter balance”, so the intelligent computer
will need to do the same, she says. Recently, Picard and her colleagues at MIT
have been designing computer vision “tools” to help video and film editors
search for specific types of images in their footage. Having learnt what, say,
houses and trees typically look like, their algorithms learn to identify similar
visual forms in other images. Importantly, they learn not just about basic
visual forms but how those forms change and vary from image to image.
But why stop there, says Picard. Why not also design search tools that can be
trained by humans to look for the most thrilling or moving or funny sequences of
a video? “If I’d heard someone suggesting doing all this a couple of years ago,
I probably would have written them off,” she says. After all, “emotions sound
like sissy science”.
Of course, designing machines to detect the external physical signs of
emotional states is not the same as creating machines that “know” anything about
emotional states, let alone ones that could be said to “feel”. “A computer may
identify a fingerprint as yours without knowing anything about fingers or how
they leave prints,” points out Aaron Sloman, professor of Artificial
Intelligence and Cognitive science at the University of Birmingham. Over the
years, Sloman has penned a string of academic papers looking at the problem of
how to simulate—rather than detect—emotional states. And his
conclusion is provocative: one day even something as complex as the inner
turmoil of grief might be modelled on computers.
Most computer whizzes and software gurus have so far been content with a view
from the sidelines. Now that’s changing. “Twenty years ago it wasn’t uncommon to
hear people talk about a new era of computers that could think,” says Michael
Casey, an MIT researcher who designs systems that respond to sound. “Today,
perhaps we’re opening up a new era of computers that can feel.”
For Casey, that means musically literate computers with an ear for the finer
points of a performance of Brahms. At MIT, he’s developed what he describes as a
“virtual accompanist”. This music-making computer system will play piano to your
violin—but not simply by retrieving a score from memory and playing it
without feeling. The computer listens to the human performer and learns to adapt
its own style. So if the human performance is brisk and light, the virtual
accompanist learns to follow suit. The machine takes its emotional “cue” from
the human performer’s timing, Casey explains in a forthcoming TV programme (
Music and the Mind, Channel 4, May 19). The system is programmed, he says,
to “warp” its own score based on the nuances of expressive timing it detects in
the human performance.
One drawback is that Casey’s virtual accompanist is for trained musicians
only. But as every spotty adolescent who has ever strummed a make-believe guitar
in the privacy of the bedroom knows, there are many more wannabes in the world
than maestros. Well, fear not. In the Californian town of Sonoma, Manfred
Clynes, neuroscientist and accomplished pianist to boot, is working on the
answer to your dreams.
Christened the Universal Musical Instrument, it’s a music-making computer
system that takes all the hard work—but not all the emotional
satisfaction—out of playing, say, concert-level piano. The computer
dutifully produces all the right notes. All you have to do is supply the musical
passion—or rather, tell the computer how you would like it to shape the
emotional quality of notes with the kinds of sublte shifts in timing and
loudness that musicians use to bring printed scores to life. Doing this
note-by-note would be tedious and difficult. But in Clynes’ system that isn’t
necessary. Instead you select a single recipe for shaping the timing and
loudness of notes within a specified time interval and apply that recipe
repeatedly until the whole score is transformed, as Clynes puts it, from an
“irritating clangor to a meaningful, living realisation”.
In the past, playing Beethoven or Bach was the preserve of a lucky few.
“Now,” says Clynes, “it becomes non-elitist to interpret music…everyone can
commune with Beethoven —without having to play a single note”. Sceptics
take heed: Clynes’ system is based on years of research showing how the shifts
in expressive timing and loudness to which music owes its emotional impact tend
to follow a cyclical pattern or “pulse” that can be measured and replicated for
individual composers.
“Emotional computing” could also transform the games scene, which is
currently dominated by characters that are in the main bellicose psychopaths. At
Carnegie Mellon University in Pittsburgh, Pennsylvania, Scott Reilly and his
colleagues are trying to put that right. They’re developing what they describe
as software tools for creating “characters with emotions and a capacity for
social interactions”. Characters, says Reilly, that real people would want to
interact with in simulated dramas.
Already Reilly has nearly a dozen rudimentary “emotional agents” to his name,
ranging from an animated house cat that gets angry when you threaten it, to an
emotionally strained gunman holding up a store. “You get to be the police
officer trying to stop him,” explains Reilly. And how the gunman reacts depends
on your tactics. Enter the store mouthing assurances that he won’t be hurt, says
Reilly, and you might succeed. “But start shooting at him and he’ll get scared,
shoot back and try to take the cashier hostage.”
In another of Reilly’s programs, you get to play the part of a child in a
playground trading baseball cards with other children. “I wanted to show that I
could design competent negotiating characters that express emotions,” says
Reilly.
The common theme here is that there’s nothing ineffable about the things that
stir our passions. And that’s a sentiment with which Sloman wholeheartedly
agrees. Joy, grief, embarrassment, shame—in the end, says Sloman, they’re
all facets of our mind’s ability to process information.
The way Sloman sees it, information about the world is constantly entering
the nervous system pulling us hither and thither. The problem is that the mind
and brain have limited resources, especially when it comes to planning and
implementing complex procedures. Hence, anyone who stumbles while carrying, say,
a tray of tea-cups is unlikely to be able to simultaneously continue with a
conversation. Mental resources must be switched instead to the more pressing
task of recovering balance. And this, by and large, is where emotional states
come in.
“We interpret the mechanisms performing such switches by saying we are in an
emotional state,” says Sloman, who believes there is no conceptual or
philosophical reason why emotional skills shouldn’t be taken apart and simulated
in computers. Still, if the painfully slow progress of past research into
artificial vision and speech is anything to go by, it could be years before
anyone knows just how to simulate realistic emotions in the mind of a
computer.
Some applications, of course, don’t require realism. Reilly’s aim, for
example, is to produce gripping drama, so his computer characters need only give
an impression of emotionality. And this he can create with a system of
behavioural goals and rules. Characters are given goals such as “find cat food”
or “escape from store with money in bag”, and rules for calculating whether, and
by how much, specific events or actions will hinder or help them achieve those
goals. The characters respond negatively, or aversively, to actions which would
conflict with their goals and positively to ones that fit those goals. And these
“feelings” influence theoutcome of the calculations that that decide how the
character will behave.
Even so, the emotional state is typically just the value of a single
variable. Is this the way to simulate realistic surges of passion, joy, fear and
rage? Is it the right type of recipe for producing the subtler feelings and
emotional “biases” that Elliott so badly lacked?
For the pursuit of new forms of dramatic art, the answer may not matter. But
Sloman sees the quest for emotional computers as a means to answer towering
questions like “Who am I?” and “How does the human mind work?”. So, for him
emotional realism is crucial.
The emotions Sloman would most like to simulate in the mind of a computer are
the complex ones—grief, humiliation, guilt, shame. Here, knowledge and
cognitive awareness are clearly vital ingredients. To feel humiliated, Sloman
points out, you need to know “who was involved, what happened, what they thought
about you, why they thought it, why you wish they hadn’t and so on”. (Sloman
suspects “it is impossible for a rat to be humiliated.”)
But such emotions are not born of knowledge and cognition alone. There’s
another vital ingredient, says Sloman—a partial loss of control over the
thoughts fuelling the emotion. Nobody consciously chooses to feel humiliated or
grief-stricken, these states happen “to” people. They involve intrusive thoughts
which people cannot easily expunge from their minds. The mind of an emotional
computer would also have to suffer such intrusions, conjectures Sloman.
But wait: aren’t computers supposed to be the very epitome of control? If
they’re in charge of flying a plane, perhaps, but not, according to Sloman, if
their job is to simulate emotional behaviour. Here, an ability to “lose it”
could be essential.
How you would make a computer system vulnerable to emotional disruptions is
not yet clear. But Sloman suggests the loss of control could emerge as a
byproduct of other features of the machine. After all, in today’s computer
networks, desirable mechanisms for memory management and time-sharing sometimes
produce undesirable states of “thrashing”—a loss of control which Sloman
likens to human panic attacks.
Damasio, as you would expect from a neurologist, looks to biology for clues
about how to simulate emotional behaviour. The way he sees it, the involuntary
or “nonconscious” side of human emotional behaviour emerges from chemical and
neural changes in the body: it’s the body pulling the strings, rather than the
conscious brain or mind. And what that means, says Damasio, is that “if we’re to
have a simulation of emotional behaviour and make claims about a machine that
would have emotions similar to those of humans, one would have to simulate a
body first”.
But whether the emotional computers of the future are based on theoretical
models of the body, brain or mind (or a combination of all three), one question
remains: are we prepared to build a machine and give up control over it? Are we
willing to give machines the freedom to make value-based, emotional
decisions?
For Picard, the path to an answer is strewn with other questions, equally
challenging but more specific. For example, should machines be given emotional
self-awareness? Should they be given emotional skills beyond the power of
humans? Should they be allowed to feign and hide emotions in the way that humans
do? The answers are by no means obvious: a machine tutoring a child with brain
damage may feel exasperation, but showing the exasperation might interfere with
the teaching process by upsetting the child.
Picard hesitates to be dogmatic, but if her vision of “affective computing”
does come to pass, she believes one rule may eventually become necessary: no
emotions without ethics. After all, if HAL hadn’t valued the success of the
mission above the lives of his crewmates, the tragedy could never have happened.
The horror of 2001 isn’t that HAL had emotions, says Picard. “It’s that he
didn’t have the intelligence and ethics to handle them.”