Watch and Learn
People who throw things for fun or profit—basketball and
baseball players, the guy who slings darts at the pub—should
respect the archer fish. This species blasts its insect prey out of
the sky or off a perch with a quick squirt of water that's accurate
up to two meters away (not bad for a fish that averages 25
centimeters in length).
However, these animals start with lousy aim, so they have to acquire
the skill over time. What's amazing is that archer fish don't have
to rehearse their piscine field goals: They can learn how to hit a
fast-moving, flying object by watching another fish do so, according
to a paper in the February 21 issue of the journal Current
Biology. Imagine a fan who watches a season of hoops on TV,
then picks up a basketball and sinks a 40-foot jump shot.
Stefan Schuster and his colleagues at Friedrich-Alexander University
of Erlangen-Nuremberg discovered this unusual ability when they set
out to study how archer fish hit their prey so accurately. The
investigators regularly presented small, stationary targets at
various heights to young fish, which became quite good at knocking
them down over the course of a year. However, a moving target
flummoxed them, even when it was going very slowly (five millimeters
per second, or about twice as fast as a speeding snail).
With time and extensive training (hundreds of repetitions), the
archer fish got better at hitting targets that moved horizontally at
constant speed. This feat requires a tricky bit of neural
computation that has to account for several variables: the target
speed and direction, the changing angle between observer and target
(which determines how much light is bent, or refracted, at the
air-water boundary), the time of flight and the effect of gravity.
As the fish became proficient, each increase in target height or
speed erased most of the previous gains in accuracy until the
shooter could adjust its calculations.
Based on this pattern, the investigators were surprised to find that
the ballistic solution painstakingly learned and revised for
horizontal movements could easily accommodate the addition of a
vertical component. For archer fish, it seems, movement in three
dimensions isn't any more difficult to follow than movement in two dimensions.
Nonetheless, the German team was unprepared for what happened when
they began another round of training with a group of five fish
unfamiliar with the moving-target game. This trial must have seemed
like a busted experiment at first, as the dominant member of the
group hogged all the shots and wouldn't let his subordinates line up
for even a single attempt at the target. As expected, the one
performing fish had learned its skills just fine by the end of
training. But when Schuster's team removed the varsity player, the
benchwarmers—fish that had previously failed to hit even the
lowest, slowest targets and hadn't taken a single shot during the
practice sessions—showed nearly the same ballistic prowess as
their former leader.
One possible explanation was that the nondominant fish figured out
all the angles simply by watching the target while the alpha
trained. But in a control group without a dominant fish
(experimenters played the same role by nudging the others out of
firing position during the training), the spectators performed just
like spectators, with dismal hit rates. In other words, the naive
fish had to learn by watching a successful group member, and this
observation apparently told them all they needed to know about
refraction, target speed, rise time and gravity effects.
Archer fish don't have impressive neural hardware to process all
this information; like other fish, they have a primitive
cerebrum, so the means of performing these computations must already
exist in simple nervous systems. At present, Schuster's team is
working to identify the neuroanatomical circuits, but he doesn't
expect to find distinctive features. The major goal is to "get
[our] hands on how a defined network performs in a cognitive
task," he says.
His success could be something of a breakthrough: Computational
neuroscientists have spent decades figuring out how animals learn
new motor skills, and most of their models use some kind of feedback
routine that tries to minimize error with each new
repetition—what engineers call a "closed-loop"
design. But shots from the archer fish are
"open-loop"—they don't involve continuous feedback.
"That's what I find most challenging," says Schuster,
"the fact that [the task] can be done extremely fast and with
remarkable precision without adhering to traditional closed-loop
design. We hope very much to convince researchers in robotics that
much speed can be saved by implementing predictive (or ballistic)
shortcuts in the motor guidance of autonomous robots."
Fish aren't particularly smart, but archer fish do seem to have a
fairly sophisticated cognitive ability to generalize from one set of
circumstances to another and to make predictions based on that
general knowledge. They clearly acquire this skill more efficiently
than we humans figure out how to swish a three-point shot or to
throw a wicked curveball. Can we learn from the fish? Well, Dr. J
is a Pisces. And Adrian Dantley, Charles Barkley and
Shaquille O'Neal are too. So I guess the answer is yes.