394
M. T. Whelan et al.
2.3 Two-Stage Dynamics
This model consists of two different behavioural states that define two different
sets of network dynamics, similar to a previous two-stage modelling approach
of CA3 replay dynamics [34]. The first behavioural state is defined as active
exploration: the state in which MiRo is actively searching for the hidden reward.
Under this state, it is assumed there is little to no synaptic transmission across
the network due to the effects of high acetylcholine levels [21], which has been
shown experimentally to inhibit the recurrent post-synaptic inputs in the hippocampal CA3 region [16]. To capture this effect, λ is set to 0 in Eq. 4 thus
preventing post-synaptic transmission.
The second behavioural state is defined as quiescent reward, in which MiRo
remains awake yet quiescent whilst it is at the reward point, and is the state
under which reverse replays occur. It is assumed here that acetylcholine levels
have dropped, similar to that found during slow-wave sleep states [21], thus
permitting synaptic transmission in the CA3 network. To model this effect, λ is
set to 1 in Eq. 4.
3 Results
3.1 Searching for a Hidden Reward
The model is run on the MiRo robot in a simulated open arena environment,
having a diameter of 2 m (Fig. 1) and using simulation time steps of 10ms. Model
equations are discretised using the Euler method with time steps of Δt = 10 ms
to match the simulated time steps.
1 From a random start location, MiRo is left
to freely explore its environment via a basic implementation of a random walk,
with the goal of finding a hidden reward. This is the active exploration phase,
and during this phase the network rates are driven solely by the place specific
inputs with no recurrent synaptic transmissions. There is no synaptic plasticity
implemented in this experiment, and so all weights, w ij , are fixed at a value of
1. Figures 2A and 2B show the activity of the network during active exploration.
Due to the distribution of the place-specific input, no more than 4 cells are
active at any one time, though most often this amounts to no more than 2 or 3
cells being simultaneously active. This sparse representation during exploration
provides a neural representation of space. Neurons that become active due to
the place specific input then undergo increases in intrinsic plasticity, decaying
exponentially (according to Eq. 8) when activity in the neuron drops.
Upon reaching the hidden reward location, MiRo pauses and enters the quiescent reward phase. Place specific inputs are computed using Eq. 3 and are
input into the network via pulses of 0.1 s-ON and 1.9 s-OFF. Recall that during
this phase, recurrent synaptic conductances are allowed. Due to the increase in
synaptic recurrent conductance and post-synaptic activity being scaled by the
1 Full code for the model (using Python 2.7) can be found at https://github.com/
mattdoubleu/robotic reverse replay.
M. T. Whelan et al.
2.3 Two-Stage Dynamics
This model consists of two different behavioural states that define two different
sets of network dynamics, similar to a previous two-stage modelling approach
of CA3 replay dynamics [34]. The first behavioural state is defined as active
exploration: the state in which MiRo is actively searching for the hidden reward.
Under this state, it is assumed there is little to no synaptic transmission across
the network due to the effects of high acetylcholine levels [21], which has been
shown experimentally to inhibit the recurrent post-synaptic inputs in the hippocampal CA3 region [16]. To capture this effect, λ is set to 0 in Eq. 4 thus
preventing post-synaptic transmission.
The second behavioural state is defined as quiescent reward, in which MiRo
remains awake yet quiescent whilst it is at the reward point, and is the state
under which reverse replays occur. It is assumed here that acetylcholine levels
have dropped, similar to that found during slow-wave sleep states [21], thus
permitting synaptic transmission in the CA3 network. To model this effect, λ is
set to 1 in Eq. 4.
3 Results
3.1 Searching for a Hidden Reward
The model is run on the MiRo robot in a simulated open arena environment,
having a diameter of 2 m (Fig. 1) and using simulation time steps of 10ms. Model
equations are discretised using the Euler method with time steps of Δt = 10 ms
to match the simulated time steps.
1 From a random start location, MiRo is left
to freely explore its environment via a basic implementation of a random walk,
with the goal of finding a hidden reward. This is the active exploration phase,
and during this phase the network rates are driven solely by the place specific
inputs with no recurrent synaptic transmissions. There is no synaptic plasticity
implemented in this experiment, and so all weights, w ij , are fixed at a value of
1. Figures 2A and 2B show the activity of the network during active exploration.
Due to the distribution of the place-specific input, no more than 4 cells are
active at any one time, though most often this amounts to no more than 2 or 3
cells being simultaneously active. This sparse representation during exploration
provides a neural representation of space. Neurons that become active due to
the place specific input then undergo increases in intrinsic plasticity, decaying
exponentially (according to Eq. 8) when activity in the neuron drops.
Upon reaching the hidden reward location, MiRo pauses and enters the quiescent reward phase. Place specific inputs are computed using Eq. 3 and are
input into the network via pulses of 0.1 s-ON and 1.9 s-OFF. Recall that during
this phase, recurrent synaptic conductances are allowed. Due to the increase in
synaptic recurrent conductance and post-synaptic activity being scaled by the
1 Full code for the model (using Python 2.7) can be found at https://github.com/
mattdoubleu/robotic reverse replay.