Computational Models for Generic Cortical Microcircuits - Computational Neuroscience: A Comprehensive Approach - page 585

periment no preprocessor that extracts these features, which are vital for a successful
prediction, each readout pool that carries out prediction for a particular sensor has to
extract on its own these relevant pieces of information from the raw and unfiltered
information about the recent history of sensor activities, which are still reverberating
in the recurrent circuit.
Twenty-eight further readout pools were trained in a similar unsupervised manner
(with 1000 training examples) to predict where the moving object is going to leave
the sensor field. More precisely, they predict which of the 28 sensors on the perime-
ter are going to be activated by more than 50% when the moving object leaves the 8
x 8 sensor field. This requires a prediction for a context-dependent time span into the
future that varies by 66% between instances of the task, due to the varying speeds
of moving objects. We arranged that this prediction had to be made while the object
crossedthecentralregionofthe8x8field,henceatatimewhenthecurrentposition
of the moving object provided hardly any information about the location where it
will leave the field, since all movements go through the mid area of the field. There-
fore the tasks of these 28 readout neurons require the computation of the direction
of movement of the object, and hence a computationally difficult disambiguation of
the current sensory input. We refer to the discussion of the disambiguation problem
of sequence prediction in [1] and [14]. The former article discusses difficulties of
disambiguation of movement prediction that arise already if one has just pointwise
objects moving at a fixed speed, and just 2 of their possible trajectories cross. Ob-
viously the disambiguation problem is substantially more severe in our case, since
a virtually unlimited number of trajectories (see Figure 18.7) of different extended
objects, moving at different speeds, crosses in the mid area of the sensor field. The
disambiguation is provided in our case simply through the context established inside
the recurrent circuit through the traces (or reverberations ) left by preceding sensor
activations. Figure 18.6 shows in panel a a typical current position of the moving
ball, as well as the sensors on the perimeter that are going to be active by

≥

50%
when the object will finally leave the sensory field. In panel b the predictions of the
corresponding 28 readout neurons (at the time when the object crosses the mid-area
of the sensory field) is also indicated (striped squares). The prediction performance
of these 28 readout neurons was evaluated as follows. We considered for each move-
ment the line from that point on the opposite part of the perimeter, where the center
of the ball had entered the sensory field, to the midpoint of the group of those sen-
sors on the perimeter that were activated when the ball left the sensory field (dashed
line). We compared this line with the line that started at the same point, but went
to the midpoint of those sensor positions which were predicted by the 28 readout
neurons to be activated when the ball left the sensory field (solid line). The angle
between these two lines had an average value of 4.9 degrees for 100 randomly drawn
novel test movements of the ball (each with an independently drawn trajectory and
speed). Another readout pool was independently trained in a supervised manner
to classify the moving object (ball or bar). It had an error of 0% on 300 test ex-
amples of moving objects. The other readout pool that was trained in a supervised
manner to estimate the speed of the moving bars and balls, which ranged from 30
to 50 units per second, made an average error of 1.48 units per second on 300 test