A Computational Approach to Memory Deficits in Schizophrenia

L.M. Talamini¹, M. Meeter¹, B. Elvevaag², T.E. Goldberg² and J. Murre¹. ¹ Dept. of Psychonomics, Univ. of Amsterdam; ²NIMH, Bethesda.

Introduction

Memory impairment is one of the most reliable and well documented neuropsychological findings in schizophrenia. Review of the available data shows notable impairments in free recall, a lesser deficit in cued recall tasks and relatively spared recognition. These data (Table 1) suggests a mild to moderate deficit in encoding and a moderate to severe deficit in retrieval of episodic memory. A less extensive body of evidence suggests normal forgetting on the short-intermediate term (minutes/hours), and abnormal recall of remote memories (months/years). This distinguishes the memory profile from the amnesia's (e.g. in Alzheimer's disease and hippocampal lesion), in which a highly increased forgetting on the short-intermediate term is observed.

The memory impairments appear to be a stable trait of schizophrenia (Table 3). Moreover, their occurrence does not show dependence on attention or other 'executive' components of learning, as seen in 'frontal patients' (Table 2). In view of the basic and stable nature of the memory impairments, we suggest that they may related to developmental abnormalities of the medial temporal lobe (MTL), which have been observed in schizophrenia. However, it is far from clear how abnormalities in the interconnected regions composing the MTL may impair encoding and retrieval without notably affecting forgetting rate.

We suggest that the the MTL circuitry subserved at least two functions, with respect to episodic memory: On the one hand fast, auto-associative learning, which is achieved through highly plastic and extensive recurrent connectivity. On the other hand the network subserves an integrative function: it needs to integrate highly convergent sensory inputs in a way that sustains not only efficient encoding, but also recall of episodic memories using cues.

In our view, the specific memory deficits in schizophrenia may be caused, not by abnormalities in the auto-associative components of the circuitry, which would be expected to result in an increased forgetting rate, but rather by an abnormality in the integrative function. This latter function might be subserved in large part by the parahippocampal region. We have used connectionist simulations to investigate this notion. The preliminary results presented here involve list learning, in a model which simulates how 'context' may be used to store and recall episodic memories (item-context associations). We propose a mechanism for integrated representation of 'context' and 'object' information and test the hypothesis that a 'hypoplasia' of the integrative module in the circuit results in recall deficits.

TABLE 1. Memory Profile in Schizophrenia
========================================
PROCES                IMPAIRMENT         INTERPRETATION

Encoding              mild - moderate    impaired entry into hippocampus

Forgetting rate       normal             normal hippocampal forgetting
(i.e. early storage)

Retrieval             moderate - severe  impaired context-dependent
                                         reactivation of memories

Learning curve        slowed             consequence of encoding deficit

Remote memory         U-shaped curve     young adult period impaired,
                                         hip-cortical transfer impaired,
                                         long-term consolidation deficit

Short term memory     mild - moderate    hippocampus (in)dependent?
(digit span)

Procedural memory     intact


TABLE 2. Dependence of Memory Deficits on Executive Task Aspects
================================================================
			      Schizophrenia 	Frontal lesion	
Attention			no 			yes
Interference			no			yes
False recognition		no			yes
Deep versus shallow encoding	no			yes
Modality			probably not 		no?


TABLE 3. Memory Impairments in Schizophrenia Correlate With:
============================================================
Outcome				yes
Negative symptoms		small correlation
Positive symptoms		no
Age				no (although old age not incl in studies)
Duration of illness		no
Medication			no

The model architecture

INPUT MODULES: We assume that sensory input reaches the hippocampus via two, more or less distinct, pathways; one processing meaningful objects, and one involved in detecting non-object visual information and spatial relations. In the model this is simulated by incorporating two input modules termed ' Item input' and 'Context input' respectively. Both have widespread, random connections to the Paralink module.

PARALINK: This module receives widespread projections from the input layers and a full projection from the Link layer. A global inhibition parameter maintains a constant activity in the layer. The 'k' neurons that are active at any given time are those that receive the highest input from the three surrounding structures. Paralink, interposed between Link and the in- and output layers, transiently holds representations of item-context associations and communicates these patterns up and down in the modular hierarchy. It is biased, by feedback from Link, towards representation of previously viewed patterns.

LINK: This auto-associative module is reciprocally connected with Paralink. It has highly plastic connections with Paralink and with itself (recurrent connections). A global inhibition parameter maintains a constant activity in the layer.

OUTPUT: Paralink nodes send a full, normalised, projection back to the Item input layer. This connection learns the relation between Paralink patterns and item patterns. The summed input to the collective nodes of an item pattern constitutes the output of the model.

Fig. 1. Input intersections in Paralink
Crosses denote Paralink nodes receiving inputs from the active context. Red nodes receive inputs from the active item nodes. The pattern connected with the item-context combinations consist of those Paralink nodes that receive inputs from both item and context (red nodes with a white cross).

Modelling learning

We simulate a list-learning experiment by presenting the model with a series of items while one context representation is activated. To represent as many unique input combinations as possible, we adopt widespread random projections from the input layers. Nodes in the intersection of the projections of active inputs (those that receive both item and context inputs) will be activated (Figure 1).

During learning the recurrent Link connections and those from Link to Paralink are modulated so that transmission is low. Hence, activity in Link is determined mostly by the input from Paralink, and activity in Paralink by the inputs. The model is allowed to learn the emerging activation patterns in Paralink and Link during a number of iterations.

Modelling free recall

Following the learning session, free recall is tested by activating the context representation and analysing the feedback activation of items by Paralink nodes. As a check on the specificity of item-context associations we then activate an alternative context and test free recall again.

During retrieval, transmission in the recurrent connections in Link is normal, changing the dynamics of Link to those of an auto-associative network in which patterns, that are partly activated by input from Paralink, can be completed. The connections from Link to Paralink enable Link to, in turn, impose a retrieved pattern on Paralink. If feedback activity to the item nodes crosses a criterion (75% of maximum), the pattern is counted as retrieved. Upon retrieval, both Link and Paralink nodes are reset. The model is allowed to cycle for 100 iterations. If, during those 100 iterations, a pattern is once retrieved to criterion, it is counted as recalled. Figure 2 shows a typical run of a free recall test.

Fig. 2. One run of the recall test
Output to the item nodes of 5 stored items, during a free recall attempt with the right context node active. Each time the output to a particular item reaches criterion value (0.75), the model is reset.

Modelling schizophrenia

A developmental hypoplasia is simulated reducing the Paralink nodes by 50%. Taking into account possible effects of hypoplasia in terms of afferent innervation and global inhibition of Paralink, the effects of node reduction on free recall are tested under three conditions:

A) Insufficient inhibition (inhibition is not scaled down to the smaller size of Paralink, leading to hyperactivation in Paralink)
B) Hyperinnervation of Paralink (number of input projections to Paralink is not scaled down to the smaller size of Paralink)
C) Both of the above

All these conditions led to lower recall performance compared to the normal situation (Figure 3), though recall suffered especially in the conditions with insufficient inhibition (A and C).

Conditions with hyperinnervation (B and C), also produced recall of items when the model was tested with the wrong context. This suggested that hyperinnervation leads to a loss of context specificity.

Fig. 3. Results of the hypoplasia simulations
Three hypoplastic conditions were compared to a control simulation in which there was no hypoplasia. The figures show the mean proportion of recalled items, out of a list of 5 stored ones, averaged over 7 replications. 'Correct recall' refers to free recall in the correct context (i.e. the one in which items were learned). 'False context' refers to recall in an alternative context, and is a measure of how much the model confuses context.

Preliminary model analysis / conclusions

The differential temporal variability of the inputs (i.e. the context input is much more stable than the item input) poses the following constraint on connectivity: the more stable input component requires a very low output variance. If not, the formation of unique representations of the co-occurrence of inputs will be comprimised, as the variability in context inputs will bias the representation of all items toward those nodes that receive the strongest context input (data not shown).

The context input also requires a denser projection (larger number of target nodes) in order to accomodate unique intersection patterns with multiple items.

We hypothesise that the item inputs on different nodes should be variable, to enable the network to activate intersection patterns of variable stringency, by varying global inhibition.

Hyperinnervation of Paralink nodes results in extensive overlap of the context inputs to Paralink. Hence, during free recall the context-cue will trigger appropriate patterns, but also patterns learned in other contexts. Notably, in this situation unique intersections in Paralink and appropriate patterns in Link have probably been learned. However, they can not be recalled via the context-cue alone. Simulated recall with part-word cues is expected to greatly improve recall in this setting.

Global inhibition determines the number of active nodes in Paralink. The amount of overlap of Paralink patterns that are learned in a given context, is proportional to the ratio between the activity level in the layer and the size of the projection field of the context nodes. The bigger this ratio, the larger the overlap. Overlap of Paralink patterns presumably also increases the overlap in Link patterns. During recall spurious activation states can occur, which do not result in item recall. Notably, with insufficient global inhibition item-context co-occurrences are not properly encoded. It might be predicted that part-word cues will not disproportionally aid recall.

Prospectus

Simulations of Cued recall and Recognition

Interference tests in a learning paradigm with two lists.

Paired associate learning

Balance between Link and Input projections to Paralink.

Use of biologically plausible neurons to investigate whether hyperinnervation of Paralink can lead to 'unstable' patterns, and thus to deficient encoding.