Technical information about datasets provided for the fly behavior modeling project

The datasets consist of sets of walking fly trajectories acquired using the setup and experimental
protocol described in Ref. 1 figure 1) A-D and supp. figure 1.
Briefly, the visual stimulus consists of repeatedly interleaved short epochs of coherent dot motion
in a particular direction (which we will call ‘motion pulses’) and random dot motion in all directions
(the ‘noise’ epoch) as shown in Fig. 1)c. In our datasets, the motion (or dark, see below) pulses lasted
150ms or approximately 5 frames.

You will receive datasets corresponding to 3 different types of experiments:
2 motion pulse presentation experiments - one where sparse dark dots where presented over a bright
background (as described on the right side of Fig. 1D and in supp. video 1) (the ‘decrement’ condition);
another where bright dots were presented over a dark background (the ‘increment’ condition). In a
3rd experiment, the flies were presented with complete darkness (‘dark pulse’) when the stimulus was
applied, and with a medium light level between stimulus applications (replacing the ‘noise epoch’).

The trajectories have been divided according to the type of flies and the type of experiment in which
they were acquired. The fly and experiment type are indicated in the file name of the corresponding
dataset as follows:

<FLY_TYPE>_<EXP>_#.mat = A matlab data file containing a structure that consists of behavioral data
(see details below); where # is a number.
<FLY_TYPE> can currently have 1 of 2 values:
control = normal flies, presenting wild-type behavior.
L2 = flies where a neuron called L2, whose functional importance for motion vision has been
demonstrated, has been inhibited.
Additional datasets will be provided at a slightly later time point to allow for a more comprehensive
performance measurement, but only the control dataset is required for parameter learning purposes.
<EXP> can have 1 of 3 values: ‘inc’ = motion pulse stimulus, increment condition. ‘dec’ = motion pulse
stimulus, decrement condition. ‘dark’ = dark pulse stimulus.

i.e., you will be getting a total of 6 structures. Within each structure, there are 11 fields containing
information of potential interest:

pos_x = fly position along the x axis
pos_y = fly position along the y axis
pos_o = fly orientation
VT = a 1st component of the fly translational velocity, in the orientation of the fly body axis (velocity of
moving forward)
VS = a 2nd component of the fly translational velocity, in the orientation normal to the fly body axis
(velocity of moving sideways. Flies do that quite often, in particular when/before/after they are turning).
VR = fly rotational velocity
Figure 2)c describes velocity histograms as well as shows how the velocities are defined. VS is not

shown, but is orthogonal to VT.
STIM = when was a stimulus applied. Samples have 1 of 3 possible values. 0 = random, incoherent, dot
motion for motion stimulus (noise epoch), medium light presentation for dark stimulus; 1 = coherent
random dot motion in one direction; 2 = random dot motion in the opposite direction (for dark stimulus,
1 and 2 are equivalent so you can pool the data from the 2 pulse presentations together).
stim_RT = an array consisting of 2 columns. Column i (where i is either 1 or 2) – the number of frames
since the first frame of the last stimulus pulse of type i that was presented. The value of stim_RT at the
first frame of a stimulus pulse is 0 and at the last frame, 4. The maximal value in stim_RT is 180 but it
is not always reached (you can either ignore these frames or assume that the new stimulus pulse was
already on – i.e. these frames are equivalent to frame 0).

Due to the specifics of the experimental setup, the trajectories can start and end at arbitrary time points
with respect to the stimulus presentation (trajectories are cut when flies step outside the ‘allowed’
region of the tube where the surface is approximately flat and they can properly see the stimulus, when
2 flies get too close to each other so that tracking fails and when flies jump). The above described fields
are all single column vectors (with the exception of stim_RT which is a 2-column matrix) consisting of
information from multiple trajectories. In order to separate the information from different trajectories,
you will use the following -

indices = a cell array. Each entry corresponds to a single fly trajectory and consists of indices into each
of the 8 fields described above. For example, pos_x(indices{3}) = the positions at each frame of the 3rd
trajectory.

Additional information (these variables should not be a part of the model, but are given to you for
complete disclosure, so that you can check whether omitting this information harms performance in any
way):

tubes = each trajectory belongs to a fly that walked in one of 7 possible tubes. This is the number of the
tube that the fly walked in. This number should not affect the fly behavior and needs not be a part of the
model you design.
day_times = different experiments were conducted at different times of the day. This field can have 3
different values: ‘a’ = morning; ‘p’ = after noon; ‘n’ = night. As with the tube number, the time of day in
which the experiment was performed should not affect the fly behavior and needs not be a part of the
model you design.

Domain knowledge:
• The direction of motion with respect to the orientation of the fly strongly affects the fly’s response
to the motion.
• The flies response to a stimulus may last longer than the duration of the presentation of the
stimulus. It is possible that the flies respond to the initiation of the stimulus as well as to its ending;
in addition to responding to the stimulus for the duration of its presentation.
• It is known that the rotational and translational velocities of flies are affected by the above
described stimuli. According to Ref. 1 these velocities are independently controlled. Can you support

/ refute this claim based on the data you are provided with?
In designing the experimental setup care was taken to make sure that the position of flies within
tubes does not affect their behavior (i.e., the fly velocities are assumed to be independent of their
position). This may not be an accurate assumption, but feel free to use it and/or experiment with
removing it.

•

References

Information about the experimental setup and expected behavioral responses can be found in:
A. Katsov and T.R. Clandinin, Motion processing streams in Drosophila are behaviorally
specialized, Neuron 59 (2008), pp. 322335. link to the paper in Science Direct
See, in particular, figure 1 panels A-D and figure 2 panels C-I.