Retrieval-Based Model Accounts for Striking Profile of Episodic Memory and Generalization

Abstract

A fundamental theoretical tension exists between the role of the hippocampus in generalizing across a set of related episodes, and in supporting memory for individual episodes. Whilst the former requires an appreciation of the commonalities across episodes, the latter emphasizes the representation of the specifics of individual experiences. We developed a novel version of the hippocampal-dependent paired associate inference (PAI) paradigm, which afforded us the unique opportunity to investigate the relationship between episodic memory and generalization in parallel. Across four experiments, we provide surprising evidence that the overlap between object pairs in the PAI paradigm results in a marked loss of episodic memory. Critically, however, we demonstrate that superior generalization ability was associated with stronger episodic memory. Through computational simulations we show that this striking profile of behavioral findings is best accounted for by a mechanism by which generalization occurs at the point of retrieval, through the recombination of related episodes on the fly. Taken together, our study offers new insights into the intricate relationship between episodic memory and generalization, and constrains theories of the mechanisms by which the hippocampus supports generalization.

Figure 1. Experimental Design.

During the study phase (top left panel) overlapping object pairs derived from triplets are presented (e.g. A1-B1, B1-C1). The test phase consists of two types of trials (top right panel): choice test trials: illustrated is a BC test trial, where participants should choose the target object (C1 (helmet) in this example), that is (indirectly) associated with the cue object (B1: clock). Source test trials: following each choice test trial, participants are asked whether the association between the cue (e.g. B1) and chosen object (e.g. C1) is “direct” (meaning that the two objects were actually presented as a pair during the study phase) or “indirect” (i.e. that the two objects were not presented together, but are indirectly associated via an unseen item). The correct response in a BC source trial would be “direct”, and in an AC trial “indirect”. Critically, source trials targeted participant’s episodic memory, for example testing whether they knew that B1 and C1 objects had actually been presented together. In contrast, participants’ responses during choice test trials probed their memory for whether objects were associated with one another (i.e. part of the same triplet), regardless of whether they had been presented together (e.g. B1-C1) or not (e.g. A1-C1). Bottom Panel: the three types of choice test trials (i.e. AB, BC, AC) are illustrated, each of which was followed by a source judgement (i.e. as depicted in top right panel). Object images were obtained from the Bank of Standardized Stimuli (BOSS) and are licensed under the Attribution-ShareAlike 3.0 Unported license (http://creativecommons.org/licenses/by-sa/3.0/).

Figure 2

Performance on experiment 1 (a), experiment 2 (b) and experiment 3 (c) for choice trials and source trials (upper and lower panels respectively). Error bars indicate the 95% highest density interval (HDI) that contains the most credible 95% of the values.

Figure 3

Reaction times for experiment 1 (a), experiment 2 (b) and experiment 3 (c) for choice trials. Error bars indicate the 95% highest density interval (HDI) that contains the most credible 95% of the values.

Figure 4. Odds ratios for AB source and BC source predictors for experiment 1, experiment 2 and experiment 3.

The odds ratios represent the odds that participants will make a correct inference judgment (i.e. correct answer on AC choice trial) given correct performance on an AB (or BC) source trial (cf. incorrect performance on respective source trial).

Figure 5. Schematic of computational models used to simulate performance in the PAI task: 2 triplets illustrated (A₁B₁C₁, A₂B₂C₂).

REMERGE: recurrent excitatory connections between feature and conjunctive layers. Curved arrow indicates competitive inhibition (i.e. softmax function) in conjunctive layer in all models. Dashed lines indicate weaker weights (cf. solid lines), implementing the assumption of weaker encoding of BC, as compared to AB, object pairs. Blend_1: the simplest model incorporating the notion of the formation of blended/integrated representations during encoding (i.e. coded by a single ABC unit). All Blend models have feedforward feature-to-conjunctive excitatory connections only. Blend_2: in addition to Blend_1, an extra AB and BC unit are included. Blend_3: in addition to Blend_2, a weight asymmetry is present denoted by dashed lines (i.e. as in REMERGE) to denote weaker encoding of BC pairs. See Methods section for details.

Figure 6

Illustration of fit of models to empirical data for choice (a) and source test trials (b) from Experiment 1. The REMERGE model provides a good qualitative fit to the profile of empirical data, and the best quantitative fit (see Table 1). See Fig. 5 and section on model architectures for details of model specification.