Spatio-Temporal Memory for Navigation in a Mushroom Body Model
421
Fig. 5. MBON membrane potential and spiking rate plotted against travelled distance.
Pink: learned route segment. Blue: unlearned segment (Color figure online).
After learning the visual flow over some travelled distance, the network memory
capacity will be reached. This is because the longer the segment, the more KCKC inhibition will be strengthened, and hence the MBON output will tend to
reduce across all input patterns, trained or untrained. This can be seen in Fig. 5
‘2.0 m’, where the low rate of MBON spiking overall could make it difficult to
distinguish learned (0–2 m) from novel (2–4 m) visual input. We return to this
issue of memory capacity in the discussion. Also note that the drop in MBON
input is less dramatic when the learned segment was taken from the middle of
the route (Fig. 5 ‘0.4 m (2)’). This reflects the effect of ‘internal noise’ in the
network dynamics: in the test, the network state at the start of the learned segment (1.2 m) depends on the preceding input, and hence differs from the initial
(resting) state used when training with that segment. This affects the relative
timing of the subsequent KC spikes, despite the same input during the segment,
so the KC pattern is not identical. Nevertheless the segment that was learned
can be easily distinguished.
Experiment 2: Adding Noise. Experiment 1 showed how the ‘internal noise’
can affect the familiarity. Here we explicitly test the effect of adding external
noise to the input, by replacing some camera events with randomly generated
events (i.e. in different pixels) at the original event time. We introduce a novelty
index P (P =
M BON Spiking Rate af ter learning
M BON Spiking Rate bef ore learning
× 100%) as a measure of the drop
in MBON activity across a segment of the route. The results in Fig. 6 show that
a noise level of 5% or more (equivalent to changing 1 input event every 1 ms)
results in output for the learned segment that can no longer be distinguished
421
Fig. 5. MBON membrane potential and spiking rate plotted against travelled distance.
Pink: learned route segment. Blue: unlearned segment (Color figure online).
After learning the visual flow over some travelled distance, the network memory
capacity will be reached. This is because the longer the segment, the more KCKC inhibition will be strengthened, and hence the MBON output will tend to
reduce across all input patterns, trained or untrained. This can be seen in Fig. 5
‘2.0 m’, where the low rate of MBON spiking overall could make it difficult to
distinguish learned (0–2 m) from novel (2–4 m) visual input. We return to this
issue of memory capacity in the discussion. Also note that the drop in MBON
input is less dramatic when the learned segment was taken from the middle of
the route (Fig. 5 ‘0.4 m (2)’). This reflects the effect of ‘internal noise’ in the
network dynamics: in the test, the network state at the start of the learned segment (1.2 m) depends on the preceding input, and hence differs from the initial
(resting) state used when training with that segment. This affects the relative
timing of the subsequent KC spikes, despite the same input during the segment,
so the KC pattern is not identical. Nevertheless the segment that was learned
can be easily distinguished.
Experiment 2: Adding Noise. Experiment 1 showed how the ‘internal noise’
can affect the familiarity. Here we explicitly test the effect of adding external
noise to the input, by replacing some camera events with randomly generated
events (i.e. in different pixels) at the original event time. We introduce a novelty
index P (P =
M BON Spiking Rate af ter learning
M BON Spiking Rate bef ore learning
× 100%) as a measure of the drop
in MBON activity across a segment of the route. The results in Fig. 6 show that
a noise level of 5% or more (equivalent to changing 1 input event every 1 ms)
results in output for the learned segment that can no longer be distinguished