快猫短视频

鈥淒O birds consider whether the nest they have built is better than the one
built in the next tree?鈥� asks Aaron Sloman, professor of artificial intelligence
and cognitive science at the University of Birmingham. 鈥淚 doubt it.鈥� If his
assessment is correct, it鈥檚 a safe bet that you won鈥檛 find a bird that feels
pride.

Sloman鈥檚 point is that being proud isn鈥檛 easy. At the very least you need a
sense of 鈥渟elf鈥� and a way to compare yourself with others. So feeling superior
takes a higher level of mental complexity than most animals can muster. That
doesn鈥檛 apply to all emotions. The fear that keeps animals out of the path of a
predator or a speeding car, for instance, is almost universal across species and
needs little or no thought.

But it鈥檚 the complex emotions like pride that fascinate Sloman. He believes
that they arise naturally from the information-processing structures that bestow
intelligence on humans, and perhaps on our nearest animal relatives. What鈥檚
more, he has a way to test his theory. Sloman has built a conceptual model of
the mind鈥檚 inner workings and is exploring ways to install it in a software
agent, a 鈥渘urse鈥� that tends children in a virtual nursery. If his ideas are
correct, his agent should begin to show complex emotions. When placed in a
society of agents, it should even have pride.

Sloman鈥檚 model resembles a cake with three layers. The lowest is a 鈥渞eactive鈥�
layer, which deals with automatic processes, low-level sensory skills such as
seeing and motor skills such as walking
(see Diagram). This is the seat of basic
emotions, such as fear and the disgust that makes you spit out bad-tasting food,
which are necessary for survival. They are probably evolutionarily old and
shared with many other species. Parts of this layer, Sloman says, resemble the
limbic system in the brain.

Forward planning

Above this sit two 鈥渕odern鈥� layers whose tasks are carried out in the cortex.
The first is a layer that manages the drives that vie for our attention. The
desire to read this article may be dominating now, but a hunger pang may well up
from the reactive layer to compete with the memory that the dog needs walking.
The 鈥渕anagement layer鈥� prioritises these drives and plans how to carry them out.
Emotions such as relief and apprehension would not exist but for our ability to
imagine different versions of events, which is a key function of the management
layer.

The final layer, Sloman calls the metamanagement layer and its job is to
manage the other two layers. This allows humans to control鈥攖o some
extent鈥攖he way they think and their emotional responses. Without this top
layer, humans would not be prone to emotions such as grief, shame, humiliation
and, of course, pride.

Emotions fit into this model in two ways. Basic emotions, such as primitive
fear, probably have dedicated, hard-wired circuits to other parts of the brain.
One word from the reactive layer鈥檚 alarm system and the rest of the brain takes
immediate notice鈥攐r the animal dies. By contrast, complex emotions such as
pride simply emerge from interactions between the model鈥檚 subsystems. Think
about a computer controlling a network. If too many people log on, the computer
can spend so much time checking who is online that it can鈥檛 do anything else.
There is no program for this behaviour, it鈥檚 just a state that emerges from the
computer鈥檚 innards when great demands are made of them.

The first hint of such behaviour is already apparent in the initial version
of Sloman鈥檚 virtual nursery. In this surreal world, the nurse is a disembodied
hand, and the babies as dots with numbers attached to them. The numbers are the
infants鈥� energy reserves and if they fall too low, the nurse must take them for
a recharge. The nurse must also stop the children falling into 鈥渄itches鈥�. If a
child does fall in, the agent must take it to a disposal point. The nurse can
sense the location and state of the children and, using Sloman鈥檚 model, must
keep them safe.

When too many children arrive at the nursery, the nurse gets 鈥渆motional鈥�. It
can become so torn between two equally urgent jobs鈥攕uch as saving one
child near a ditch and taking another to be recharged鈥攖hat it cannot carry
out either. It鈥檚 almost as though the agent is hassled. But, of course, it can鈥檛
鈥渇eel鈥� hassled. 鈥淭hat requires the metamanagement layer to detect the state,鈥�
says Sloman.

Though his nurse-agent is at an early stage, Sloman is hopeful that he can
put it in touch with more sophisticated emotions such as pride. To have pride in
its work an agent would need to know that its job can be done well or badly and
be able to assess its own performance against that yardstick. Stepping up a
gear, an agent capable of learning could use this assessment to change the way
its management layer functions, helping it to continue improving. 鈥淎 further
level of sophistication would be the ability to remember the evaluation of
previous performances and the ability to detect whether things are getting
better or worse,鈥� says Sloman.

Among Sloman鈥檚 goals is to build social agents that can compare their
performance with others鈥�. 鈥淔or an agent which has that level of conceptual
sophistication, there may be a recognition that others in the group rank it
highly,鈥� says Sloman. To be even more human-like, the agent would also need the
capacity to find the state associated with pride pleasant.

This is where pride as 鈥渟in鈥� could creep in. If the self-evaluation went
awry, an agent might, say, feel itself to be better than others. This overblown
self-esteem might arise because the agent really was better at its task than
others and took pleasure in pointing it out. Or it could become just part of the
agent鈥檚 attitude, says Sloman. People can become proud because a belief in their
superiority is reinforced by family, culture or religion. In an agent, such
reinforcement would bias the complete workings of the model of mind.

Sloman has similar 鈥渟hopping lists鈥� for other complex emotions, most of them
in the same conceptual form. 鈥淭he concepts are inherently indeterminate,鈥� he
says, 鈥渓ike the concepts of `water鈥� and `air鈥� prior to the development of the
atomic theory of matter.鈥� Nonetheless, models of mind can help to improve our
understanding of these concepts, he says. AI models may share design features
with the brain and give insights into how it works.

In recent years, some AI researchers have abandoned hierarchical structure
like Sloman鈥檚 layers. They have built robots, for example, by simply linking a
sensor to a wheel motor via a neural network and given them a small collection
of simple rules to follow. These robots can show complex behaviours that
resemble emotions like aggression. Rosalind Picard of the Media Lab at the
Massachusetts Institute of Technology believes that Sloman鈥檚 approach has the
edge over these simpler techniques. To separate basic from complex emotions, she
says, 鈥測ou鈥檙e going to need the structure at some level鈥�.

Only further versions of Sloman鈥檚 agents will reveal whether his structures
are anything like those that operate within the human mind. If he comes close,
that at least will be something to be proud of.