Improving Recall in an Associative Neural Network Model of the Hippocampus
9
Fig. 6. Mean recall quality of ‘model 1’, ‘model 2’, ‘model 3’ and ‘model 2010’ as a function of
number of stored patterns. Variable number of ‘active cells per pattern’ (5, 10, 20) in a network
of 100 PCs with 40% overlap.
increased as more and more patterns were stored in the network, while keeping the
BSC firing rate constant. In all simulations, ‘model 1’ outperformed ‘model 3’ across
all conditions (overlaps and ‘active pattern cells’). This was due to the fact in ‘model
1’ BSC was excited by 100 CA3-PCs at high frequency (40 Hz), whereas in ‘model 3’
BSC was excited by 20 CA1-PCs that fired once or twice. Since in ‘model 1’ the BSC
firing frequency response is higher than in ‘model 3’, then the postsynaptic effect of
BSC on the PC dendrites in ‘model 1’ is higher in frequency and duration (but not in
amplitude) than in ‘model 3’ (see Fig. 4A & C). Thus, ‘model 1’ has a better success at
removing spurious activities and improving recall quality than ‘model 3’. Since the BSC
frequency response in ‘model 2’ was fixed, but its postsynaptic effect (weight) on PC
dendrites increased, then the amplitude of the inhibitory postsynaptic potentials (IPSPs)
on PC dendrites increased (compared to the IPSP amplitudes in models 1 and 3), but
their frequency response was low (lower than in models 1 and 3; see Fig. 4B). Each
IPSP decayed to almost zero before another IPSP was generated post-synaptically on
PC dendrites.
4 Conclusions
A bio-inspired neural model of hippocampal CA1 region [1] was employed to systematically evaluate its mean recall quality against number of stored patterns, percent overlaps
and ‘active cells per pattern’. Modulation of selective excitatory and inhibitory pathways
to BSC and PC dendrites as more and more patterns were stored in the network of 100
CA1-PCs resulted into three models, the performances of which were compared to that
of the original 2010 model of Cutsuridis et al. [1]. Of the three models tested, ‘Model’s
9
Fig. 6. Mean recall quality of ‘model 1’, ‘model 2’, ‘model 3’ and ‘model 2010’ as a function of
number of stored patterns. Variable number of ‘active cells per pattern’ (5, 10, 20) in a network
of 100 PCs with 40% overlap.
increased as more and more patterns were stored in the network, while keeping the
BSC firing rate constant. In all simulations, ‘model 1’ outperformed ‘model 3’ across
all conditions (overlaps and ‘active pattern cells’). This was due to the fact in ‘model
1’ BSC was excited by 100 CA3-PCs at high frequency (40 Hz), whereas in ‘model 3’
BSC was excited by 20 CA1-PCs that fired once or twice. Since in ‘model 1’ the BSC
firing frequency response is higher than in ‘model 3’, then the postsynaptic effect of
BSC on the PC dendrites in ‘model 1’ is higher in frequency and duration (but not in
amplitude) than in ‘model 3’ (see Fig. 4A & C). Thus, ‘model 1’ has a better success at
removing spurious activities and improving recall quality than ‘model 3’. Since the BSC
frequency response in ‘model 2’ was fixed, but its postsynaptic effect (weight) on PC
dendrites increased, then the amplitude of the inhibitory postsynaptic potentials (IPSPs)
on PC dendrites increased (compared to the IPSP amplitudes in models 1 and 3), but
their frequency response was low (lower than in models 1 and 3; see Fig. 4B). Each
IPSP decayed to almost zero before another IPSP was generated post-synaptically on
PC dendrites.
4 Conclusions
A bio-inspired neural model of hippocampal CA1 region [1] was employed to systematically evaluate its mean recall quality against number of stored patterns, percent overlaps
and ‘active cells per pattern’. Modulation of selective excitatory and inhibitory pathways
to BSC and PC dendrites as more and more patterns were stored in the network of 100
CA1-PCs resulted into three models, the performances of which were compared to that
of the original 2010 model of Cutsuridis et al. [1]. Of the three models tested, ‘Model’s