Memory consolidation from seconds to weeks: a three-stage neural network model with autonomous reinstatement dynamics

Abstract

Declarative long-term memories are not created in an instant. Gradual stabilization and temporally shifting dependence of acquired declarative memories in different brain regions-called systems consolidation-can be tracked in time by lesion experiments. The observation of temporally graded retrograde amnesia (RA) following hippocampal lesions points to a gradual transfer of memory from hippocampus to neocortical long-term memory. Spontaneous reactivations of hippocampal memories, as observed in place cell reactivations during slow-wave-sleep, are supposed to drive neocortical reinstatements and facilitate this process. We propose a functional neural network implementation of these ideas and furthermore suggest an extended three-state framework that includes the prefrontal cortex (PFC). It bridges the temporal chasm between working memory percepts on the scale of seconds and consolidated long-term memory on the scale of weeks or months. We show that our three-stage model can autonomously produce the necessary stochastic reactivation dynamics for successful episodic memory consolidation. The resulting learning system is shown to exhibit classical memory effects seen in experimental studies, such as retrograde and anterograde amnesia (AA) after simulated hippocampal lesioning; furthermore the model reproduces peculiar biological findings on memory modulation, such as retrograde facilitation of memory after suppressed acquisition of new long-term memories-similar to the effects of benzodiazepines on memory.

Keywords: anterograde amnesia; complementary learning systems; computational model; memory consolidation; neural adaptation; retrograde amnesia; retrograde facilitation; synaptic depression; working memory.

Figure 1

The three-stage memory model: from prefrontal short-term memory to long-term neocortical memory. Activity in cortical areas (PFC, CTX) is organized into hypercolumns, while HIP activity is sparser, pattern-separated, and lacks columnar organization.

Figure 2

Example of learning (15 random patterns, sequentially trained over 150 ms) and subsequent autonomous replay activity in a small BCPNN (50 Units in 5 Hypercolumns). Note that some early patterns are never replayed due to forgetting, while other, stronger patterns reactivate multiple times. Longer reactivations are often a sign of less correlated patterns.

Figure 3

Activity in HIP changes with the cortical input, as measured by normalized activity overlap. The hippocampus implements a pattern separation mechanism which yields highly significant changes in activity when cortical input changes only slightly. Left: randomly varying the activity of 1–10 of the 50 cortical hypercolumns yields highly significant changes in hippocampal encoding. Note that to highlight the variability of coding, error bars denote one STD, not SEM (^***We test against H₀: median difference between the pairs is zero. As normal distributions of overlaps are not guaranteed in this case, we use the non-parametric pairwise Wilcoxon signed-rank test, yielding p < 10⁻¹⁶⁵, when using 1000 pairs.) Right: effective pattern separation can also be seen from the fact that hippocampal patterns diverge much faster than cortical input, e.g., changing the activity of just one cortical hypercolumn yields a 2% CTX pattern change (as measured by 98% overlap), but nearly 17% in HIP. When we change 10 hypercolumns (80% cortical overlap) then about half of the originally active HIP units are no longer a part of the encoded pattern. Note again that error bars denote STD, rather than SEM.

Figure 4

The simulation cycle with its three alternating phases, named perception, reflection, and sleep. Online learning occurs only during perception. All other learning is a function of memory consolidation during reflection and sleep. The gating of various projections at the transition between simulation phases is summarized also in Table 2.

Figure 5

The three simulation phases 1–3 and their active components, as well as the configuration during cued recall after many days of consolidation. During perception, feed-forward projections from neocortical input generate separate PFC and HIP traces, which are associated to the CTX trace via Hebbian-learning in the back-projections. This online learning episode is very brief and effectively too short to establish lasting HIP and CTX memories. During the reflection phase, replay in PFC (similar to active rehearsal) generated by the interplay of its auto-association and adaptation projections, drives HIP reinstatements, thus facilitating learning in its auto-associative projections. During sleep, HIP replay then drives CTX reinstatements which facilitate long-term learning. During cued recall, the external neocortical activation generates corresponding cues in PFC and HIP through feed-forward connections. All three networks are then individually or simultaneously allowed to relax/converge to attractors, potentially yielding successful recall of a corresponding training pattern. ^***It should be noted, that the strongest influence of the PFC on the hippocampus in primates is indirect through parahippocampal cortices. The direct projection PFC-to-HIP is neuroanatomically non-existent (Otani, 2004). We consider this modeling issue in the discussion.

Figure 6

Replay drives reinstatements of earlier patterns in HIP or CTX, respectively. The length of these reinstatements is distributed around an average reinstatement length of 40.08 ms for PFC-driven HIP reinstatements during reflection and 95.43 ms for HIP-driven CTX reinstatements during sleep.

Figure 7

Consolidation, as measured by recall rates of training patterns from each stage (PFC, HIP, CTX). By averaging the recall rates for patterns introduced on the same day, we obtain a more direct relationship between the recall rate and the age of a pattern in days. Combined recall from all stages (solid lines) is shown with and without hippocampus (full lesion) to illustrate its importance for patterns of different age.

Figure 8

Five different amnesia gradients. Retrograde amnesia after full hippocampal lesioning, anterograde amnesia (performance measured after using the lesioned system for 39 days) with different lesion size and persistent sleep deprivation, where we cut the length of the sleeping phase by half.

Figure 9

(A,B) Behavioral responses of animals receiving extensive hippocampal system lesions (circles) or control lesions (squares) as a function of the numbers of days elapsing between exposure to the relevant experiences and the occurrence of the lesion. Bars surrounding each data point indicate the standard error. Panel (A) shows the percentage choice of a specific sample food (out of two alternatives) by rats exposed to a conspecific that had eaten the sample food. Panel (B) shows fear (freezing) behavior shown by rats when returned to an environment in which they had experienced paired presentations of tones and footshock. Data in Panel (A) are from Winocur (1990). Data in Panel (B) are from Kim and Fanselow (1992). The added lines are from a simple differential equations fit from a previous modeling attempt (McClelland et al., 1995). Panel (C): Combined retrieval rates of the normal and hippocampally lesioned simulation model. Rather than the standard error (which is too small to show, as we average 500 simulations), error bars indicate a standard deviation of the underlying data, showing the stochasticity of the consolidation process.

Figure 10

In this experiment, we boosted hippocampal plasticity during learning of percept 89 (consisting of patterns 265–267) by a factor of two (halfing τ_L) and tested recall 5 days later. Top: consolidation curves showing the probability of successful recall 5 days after introduction of percept 89. Middle: the absolute change of recall probability vs. controls (simulation without any modulation). Bottom: the time course of consolidation for the modulated percept, as measured by testing HIP and CTX recall every day following the original learning experience.

Figure 11

In this experiment, we simulated the memory impact of triazolam with a half-life of 2 h by reducing hippocampal plasticity by a factor of 10 and decaying this modulation with a 2 h half-life to the original level of plasticity. The modulation was triggered at the introduction of percept 89. Top: consolidation curves measured 5 days after the modulation event, showing the lasting effect on the probability of successful recall. Middle: the absolute change of recall probability vs. controls. Note that the y-axis was broken to also visualize the smaller impact seen in the other, unmodulated percepts. Bottom: the time course of consolidation for the modulated percept, obtained by testing recall from HIP and CTX every day following the original learning experience.