may dictate intelligence
Technology Research News
Just how much does intelligence depend
A pair of researchers has made the somewhat surprising finding that context
makes a big difference -- at least in a simulation where tiny artificial
neural networks must choose which of two groups to join.
This fundamental behavior could underlie much more complicated systems
like the human brain, and may point to better methods of creating artificial
intelligence systems, said Joseph Wakeling, a graduate student at
the University of Fribourg in Switzerland.
In the researchers' simulations, 251 of the neural network agents evolved
through many rounds of the simple task. The agents congregated into two
groups, and those that ended up in the smaller group won. The agents decided
which group to join based on the context of past results.
The agents were designed to mimic biological brains, which all consist
of a set of inputs and outputs, a mechanism to make decisions, and a system
for determining whether a given output has been successful in providing
appropriate feedback, according to Wakeling.
In the exceedingly simple E. coli bacterium, the inputs test for glucose,
and the decision is to move in the direction of glucose if the bacterium
senses it, and otherwise move in a random direction.
The agents used sets of input neurons to remember a certain number of
past turns. Forty-eight intermediary neurons were each connected to every
input neuron via synapses of different strengths. These neurons provided
the agents' analytic capability, making decisions about what to do in
a current round based on the sum of the strength of the synapses connecting
them to past decisions.
The intermediary neurons responsible for winning decisions to join the
smaller group continued into the subsequent round unscathed, while those
that chose what ended up to be the larger, losing group were changed.
"Those agents who lose [reduced the strength] of the intermediary synapses
that were activated that turn. The idea [is] that these synapses were
responsible for a bad decision and should therefore not be used again,"
said Wakeling. The process is essentially a "Darwinian evolution of good
behavioral patterns," he said.
The point of the experiment was to find the fundamental mechanics behind
all naturally occurring neural systems, from the bacterium to the octopus
to the human, Wakeling said. "What are the basic mechanisms at work that
[are common to] the human brain, a mouse's brain, a lizard's brain?" he
said. There might be some universal, simple mechanism that allows a large
number of neurons to connect in a way that helps the organism to survive,
The simulation used a specific task to try to tease out a universal intelligence
mechanism rather than relying on specific characteristics like the number
of intermediary neurons the brain has, said Wakeling "The only truly objective
measure we [could] think of was success rate at solving some problem --
in other words, success in the context of the surrounding environment,"
As it turned out, the success rates of the agents were highly dependent
on the exact nature of the competition, leading the researchers to conclude
that intelligence is all about context, he said.
In the researchers' simulations, the agents that evolved together tended
to think alike, and therefore did not do very well at choosing the less-crowded
group. In fact, not one agent achieved even a 50-percent success rate
even when some were granted more memory than others. In this context,
they would all be better off flipping coins to make their decisions, according
When the researchers introduced a single, rogue agent that had more memory
than the others, however, the rogue agent was stunningly successful, choosing
the right group 99.8 percent of the time. "The key is that the rogue
is unique in having this extra memory," said Wakeling.
The researchers concluded that an agent can only be truly successful if
there are other agents around whose weaknesses it can exploit. If the
behavior of the others is highly unpredictable, or they are capable of
biting back, the agent's chances of success are vastly reduced.
Many real-world strategies, like making financial investments, also fail
because they're based on common knowledge, and most people's approaches
will be similar, according to Wakeling.
Because of this mechanism, the researchers could not predict if a given
agent would be good or bad at choosing the right group without knowing
about the other agents it would be competing against. This shows that
the question of how intelligent a system is can only be answered by examining
how good it is at coping with its surrounding environment, said Wakeling.
Human beings, after all, are more intelligent than other animals because
we are more successful at manipulating our environment to our own benefit,
The conclusion may also point the way to better methods of making artificial
intelligence. Designing an intelligence to operate within a certain environment
may prove more useful than creating a consciousness in a box and then
giving it a purpose, Wakeling said.
The research "takes the work on modeling competition a bit further in
an interesting way," said Frank Ritter, an associate professor of information
science and technology and psychology at Pennsylvania State University.
"Showing how different memory leads to different behavior, and how context
is important for problem solving," is interesting work, he said. The conclusions
make sense particularly in an exercise like choosing the smaller of two
groups, where the problem is about the context, he added.
It's an open question whether the effect can be generalized, however,
said Ritter. "In other architectures, and perhaps with other tasks, this
effect might not be seen."
Wakeling's research colleague was Per Bak. They published the research
in the November, 2001 issue of the journal Physical Review E.
Timeline: 5 years
TRN Categories: Applied Computing; Artificial Intelligence;
Story Type: News
Related Elements: Technical papers, "Intelligence Systems
in the Context of Surrounding Environment," Physical Review E, November,
2001; "Adaptive learning by external dynamics and negative feedback,"
Physical Review E, March, 2001.
Nerve-chip link closer
Inside-out gem channels
Computer follows video
may spot intruders
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog
Buy an ad link