Memory GAPS: Would LLMs pass the Tulving Test?

Abstract

The Tulving Test was designed to investigate memory performance in recognition and recall tasks. Its results help assess the relevance of the "Synergistic Ecphory Model" of memory and similar RK paradigms in human performance. This paper starts investigating whether the more than forty-year-old framework sheds some light on LLMs' acts of remembering.

Figures (3)

• Figure 1: The GAPS: Elements of Episodic Memory and their Relations. The element of encoding is a process that converts the information about an experienced event or episode (in a particular setting, at a particular time) into an engram or memory trace. The central element of the retrieval processes ecphoric information, a synergistic product of the engram and the retrieval cue, which calls on both episodic and semantic information. Source for figure: Ch. 7, citeciteproc_bib_item_1[1, 7-1, p. 135].
• Figure 2: The Transformer Architecture. Based on the 2017 paper citeciteproc_bib_item_5[5] attention mechanism, the Transformer architecture requires less training time than previous recurrent neural architectures. Input text is split into tokens (sometimes called n-gram, dangerously reminiscent of Semon's engrams--see text), then converted into vectors. Through different layers, each token is contextualized with other tokens via parallel attention heads, calculating weights for each according to its importance. The Transformer Architecture elaborates on softmax-based attention mechanism citeciteproc_bib_item_6[6] and Fast Weight Controllersciteciteproc_bib_item_7[7]. Source for figure: citeciteproc_bib_item_5[5].
• Figure 3: The Synergistic Ecphory Model (SEM). Source for figure: Ch.14 citeciteproc_bib_item_1[1, 14.3]. Schematic diagram depicting a given episode such as the appearance of a familiar word in a particular study list. Variations of trace information, a, b and c correspond to different engrams resulting from many different possible encodings of the same event, only some of which are realized on a particular occasion. Retrieval information x, y, z correspond to the different potentially relevant retrieval cues (recall) that may or may not be present on a particular occasion. (We simplify here under the assumption of a single dimension for memory traces and a single dimension for retrieval information, resulting in a bi-dimensional ecphoric vector space.) The curved lines represent conversion thresholds for different memory tasks, here recognition/familiarity and recall/identification. According to the position of the point representing the synergy of retrieval and trace information relative to the threshold lines, the rememberer would pass or not the given test.