Pattern association and consolidation emerges from connectivity properties between cortex and hippocampus

Abstract

The basic structure of the cortico-hippocampal system is highly conserved across mammalian species. Comparatively few hippocampal neurons can represent and address a multitude of cortical patterns, establish associations between cortical patterns and consolidate these associations in the cortex. In this study, we investigate how elementary anatomical properties in the cortex-hippocampus loop along with synaptic plasticity contribute to these functions. Specifically, we focus on the high degree of connectivity between cortex and hippocampus leading to converging and diverging forward and backward projections and heterogenous synaptic transmission delays that result from the detached location of the hippocampus and its multiple loops. We found that in a model incorporating these concepts, each cortical pattern can evoke a unique spatio-temporal spiking pattern in hippocampal neurons. This hippocampal response facilitates a reliable disambiguation of learned associations and a bridging of a time interval larger than the time window of spike-timing dependent plasticity in the cortex. Moreover, we found that repeated retrieval of a stored association leads to a compression of the interval between cue presentation and retrieval of the associated pattern from the cortex. Neither a high degree of connectivity nor heterogenous synaptic delays alone is sufficient for this behavior. We conclude that basic anatomical properties between cortex and hippocampus implement mechanisms for representing and consolidating temporal information. Since our model reveals the observed functions for a range of parameters, we suggest that these functions are robust to evolutionary changes consistent with the preserved function of the hippocampal loop across different species.

Figure 1. Neural network model and study design.

a) The network consists of a cortical network as well as a hippocampal input and output layer. It is described by the parameters c: number of input/output connections for each hippocampal neuron; d: conduction delay between cortical and hippocampal neurons; h: number of hippocampal input/output neurons. b) STDP rule of the neural network (after Fig. 4 in Izhikevich [18]). c) Different interpretations of a conduction delay between cortex and hippocampus. If a signal generated by the cortex needs a time interval of 2d to propagate to the hippocampus and back to the cortex, the interval can be represented by two different models. Either there is a delay of d between hippocampus and cortex and in the back direction, but no processing takes place in the hippocampus, or the delays amount to only a fraction of d, but hippocampal processing consume the remaining time interval. d) The design of this study: In a preparatory step, we sample the behavior of the network across the parameters d, h and c and adjust the parameters of no interest to ensure that network behavior is consistent for a range of parameters of interest. Subsequently, we analyze the functional contribution of the parameters of interest (degree of connectivity, delay) on the learning of associations between cortical patterns.

Figure 2. Behavior of the cortico-hippocampal network.

a) Spiking of networks with parameters h = 100, c = 300 and d between 20 and 120 ms. All other parameters of the model as in Izhikevich (2006). Strong oscillations can be observed with synchronous spiking of all cortical neurons. b) Scatter plots of networks in the 3-dimensional parameter space. Networks represented by black dots show strong oscillatory behavior as depicted in a). Gray dots indicate varied spiking behavior. For essentially all models with h>50, addition of the hippocampus to the network leads to overload causing these oscillations. c) Network behavior across the parameter space after the parameters of no interest have been adjusted as described in the text.

Figure 3. Two cortical patterns can be associated in two different ways.

The target pattern (pattern 2) can be activated by the cue (pattern 1) either directly by cortico-cortical connections or indirectly through the hippocampal loop. Low random activity (as shown after 120 ms) do not cause any spiking in the hippocampus. Black arrows show connections which are predefined and are not modified by STDP. Gray arrows indicate the connections that undergo STDP and therefore are involved in learning.

Figure 4. Pattern association through a hippocampal loop.

Panel a) and b) show the recall of two previously learned associations. Both recalls were initiated by an external cue at around 100 ms, which evoked a sparse temporal pattern in the hippocampal layers, which in turn drove the target pattern in the cortex. c) Average number of neurons for all four patterns during learning and recall. Grey bars indicate when patterns have been externally stimulated. d) Mean weight for connections between hippocampal neurons that spike in response to pattern A and C (H(A) and H(C), respectively) to target patterns B. Grey bars as in c). During the first learning phase, the connection weight increase and thereby contribute to the association of pattern A with B. Although in the second learning phase the same hippocampal neurons are also activated by pattern C, their connections to neurons of pattern B remain stable indicating that the spatio-temporal sequence of hippcampal neurons is important for storing unique associations.

Figure 5. Neurons in the target pattern are driven by three different mechanisms.

Thick lines mark connections, where one hippocampal output neuron can generate a spike in a cortical neuron. Thin lines indicate that these connections alone cannot generate a spike. Instead timed activity of more than one hippocampal neuron is needed to elicit a spike in the cortical neuron. Unconnected dots represent cortical neurons whose spiking is not directly evoked by the hippocampus, but by recurrent cortical inputs. Already in the learning process, autoassociative connections between the cortical neurons in target pattern emerge.

Figure 6. Influence of anatomical properties of the hippocampal loop on the capability to associate patterns across.

Each plot depicts the number of target neurons that are activated during recall as a function of the temporal separation between the cue and target patterns during training. From top to bottom: original model by Izhikevich , hippocampal loop with high connectivity/low delay, low connectivity/high delay, high connectivity/high delay, and high connectivity/high delay+temporal dispersion.

Figure 7. Temporal dispersion in the hippocampal loop.

a) Temporal dispersion between hippocampal input and output layer generates a sequence of spikes that increases the time span in which cortical patterns can be learned. b) Anatomical interpretation of the model. The delay between cortex and hippocampal loop is comparatively short. Hippocampal processing time varies between 10–90 ms.

Figure 8. Repeated recalls leads to temporal compression of pattern association.

Time separation Δt during recall vs. separation during learning, directly after learning (solid line) and after 300 recalls (dashed line). Each data point represents the average of 10 repetitions in which the median timepoint for the activation of the target pattern was recorded. With temporal dispersion, patterns with any interval of Δt during learning can be learned. When the hippocampal loop is involved in learning (Δt >70), the temporal separation between cue and target pattern directly after learning is result of Δt during learning. After 300 recalls the target pattern is activated after ∼10 ms. In the model without temporal dispersion, the target pattern remains at a temporal separation of either ∼10 ms or ∼120 ms. Asterisk and arrow indicate significant differences between the median timepoint of the target pattern directly after learning and that after 300 recalls (Mann-Whitney-Wilcoxon, p<10⁻⁴).

Figure 9. Illustration of the mechanism that compresses the temporal sequence of spikes.

a) The weight of a connection between a pre- and post-synaptic neuron increases according to the STDP rule, even if the interval between pre- and post-synaptic spike exceeds the delay of the connection. However, when the connection is strong enough, the pre-synaptic neuron can directly cause the post-synaptic neuron to fire. Thereby, the temporal interval of both spikes becomes shortened. b) The same effect applies when timed spikes of several neurons are required. In this example, the neurons a-j are connected with neuron k and fire in a temporal sequence. Neuron i and j together can make neuron k spike. However, STDP also increases connection weights from neurons spiking earlier in time (e.g. h and g). After sufficient repetitions, neuron i can directly evoke the spike of the target neuron as the contribution of neuron h already increased the activity level of neuron k. Thereby, neuron k spikes earlier. This mechanism continuous for all preceding neurons.

Figure 10. The association between cue and target patterns becomes independent of the hippocampus with repeated recalls.

a) Histogram of target neuron spikes as a function of time and recalls. At the beginning of the learning phase, neurons in the target pattern fire spikes with a delay of 120 ms. During later learning trials and recall, more and more neurons fire earlier until cortical neurons are directly linked with the target pattern. b) Plot of the spikes of the target neurons relative to the onset of the cue pattern. The upper and lower half of the plot depict the spiking pattern directly after learning and after 300 recalls, respectively (indicated by arrows in c). c) Median of the histogram in a). Due to the outliers in the recall phase the mean value would not be representative of most spike times. d) Number of neurons in the target pattern that can be evoked without the hippocampal loop, as a function of recalls.

Figure 11. Consolidation across a range of model parameters.

Consolidation is robustly observed in our model for a range of two parameters: the number of hippocampal neurons, and the number of cortical cells that each hippocampal cell is connected with. Black dots depict networks that show consolidation, networks represented by circles did not reveal consolidation. Behavior of the model across the parameter space when a) only one or b) eight associations are stored in the network.