Fast Reverse Replays in a Robotic Hippocampal Model
391
There are a number of important differences in both the behavioural and
neural states between reverse replays and forward replays. For instance, whilst
forward replays show to occur during both awake and sleep states, and can
reinstate either remotely (i.e. reinstating experiences from locations that are
spatially distant from the original experience) or locally [4], reverse replays tend
to occur almost exclusively during awake replay events, reflecting the immediate
local past experience [11]. And unlike awake forward replays, reverse replays
are strongly modulated by rewards, such that the frequency of reverse replays
initiating at reward sites increases in the presence of increased rewards and
reduces in response to decreased rewards [1]. In addition, reverse replays are
capable of emergence following only a single trial, and it has therefore been
proposed that reverse replays may be a consequence of place cell excitability
changes not relying on traditional synaptic plasticity, such as lingering place cell
activities [8], or more recently long-term potentiation of intrinsic excitability [30].
Reverse replays may then induce synaptic changes that provide the mechanism
for replays during states where the transient activity has faded, such as during
sleep states [2]. These point towards an intricate interplay between transient
excitability changes and synaptic plasticities in the emergence of the various
forms of replay under different behavioural states.
Presented here is a biophysical, continuous rate-based network model of
reverse replay implemented on a simulated version of the biomimetic robot MiRo
[25], and is based on two recent models of hippocampal replay dynamics in recurrent CA3 networks [15,30]. Reverse replays in this model occur as a consequence
of two modes of transient neural states. The first is due to the implementation
of a time decaying model of intrinsic plasticity. Intrinsic plasticity is the ability of a cell to increase heterosynaptic long-term potentiation of post-synaptic
potentials following recent activity [17,44], and has recently been proposed as
a potential mechanism for the occurrence of reverse replays [30]. The second
transient neural state implementation is in short-term plasticity, which acts to
ensure unidirectional, stable replays [15]. This is due to short-term depression
suppressing synaptic currents after a given amount of continuous firing, thus
preventing unbounded synaptic transmissions.
2 Methods
2.1 Network Architecture
The network consists of 100 rate-based neurons representing place cells, arranged
in a grid of size 10 × 10, each of which has its place fields spread evenly across
an open circular environment. Each cell forms a bidirectional and symmetric
synaptic connection to its 8 nearest neighbours, with all weights fixed at a value
of 1. Figure 1 gives an example of the network architecture for a subset of cells.
391
There are a number of important differences in both the behavioural and
neural states between reverse replays and forward replays. For instance, whilst
forward replays show to occur during both awake and sleep states, and can
reinstate either remotely (i.e. reinstating experiences from locations that are
spatially distant from the original experience) or locally [4], reverse replays tend
to occur almost exclusively during awake replay events, reflecting the immediate
local past experience [11]. And unlike awake forward replays, reverse replays
are strongly modulated by rewards, such that the frequency of reverse replays
initiating at reward sites increases in the presence of increased rewards and
reduces in response to decreased rewards [1]. In addition, reverse replays are
capable of emergence following only a single trial, and it has therefore been
proposed that reverse replays may be a consequence of place cell excitability
changes not relying on traditional synaptic plasticity, such as lingering place cell
activities [8], or more recently long-term potentiation of intrinsic excitability [30].
Reverse replays may then induce synaptic changes that provide the mechanism
for replays during states where the transient activity has faded, such as during
sleep states [2]. These point towards an intricate interplay between transient
excitability changes and synaptic plasticities in the emergence of the various
forms of replay under different behavioural states.
Presented here is a biophysical, continuous rate-based network model of
reverse replay implemented on a simulated version of the biomimetic robot MiRo
[25], and is based on two recent models of hippocampal replay dynamics in recurrent CA3 networks [15,30]. Reverse replays in this model occur as a consequence
of two modes of transient neural states. The first is due to the implementation
of a time decaying model of intrinsic plasticity. Intrinsic plasticity is the ability of a cell to increase heterosynaptic long-term potentiation of post-synaptic
potentials following recent activity [17,44], and has recently been proposed as
a potential mechanism for the occurrence of reverse replays [30]. The second
transient neural state implementation is in short-term plasticity, which acts to
ensure unidirectional, stable replays [15]. This is due to short-term depression
suppressing synaptic currents after a given amount of continuous firing, thus
preventing unbounded synaptic transmissions.
2 Methods
2.1 Network Architecture
The network consists of 100 rate-based neurons representing place cells, arranged
in a grid of size 10 × 10, each of which has its place fields spread evenly across
an open circular environment. Each cell forms a bidirectional and symmetric
synaptic connection to its 8 nearest neighbours, with all weights fixed at a value
of 1. Figure 1 gives an example of the network architecture for a subset of cells.