How do we generalize?

Abstract

Humans and animals are able to generalize or transfer information from previous experience so that they can behave appropriately in novel situations. What mechanisms–computations, representations, and neural systems–give rise to this remarkable ability? The members of this Generative Adversarial Collaboration (GAC) come from a range of academic backgrounds but are all interested in uncovering the mechanisms of generalization. We started out this GAC with the aim of arbitrating between two alternative conceptual accounts: (1) generalization stems from integration of multiple experiences into summary representations that reflect generalized knowledge, and (2) generalization is computed on-the-fly using separately stored individual memories. Across the course of this collaboration, we found that–despite using different terminology and techniques, and although some of our specific papers may provide evidence one way or the other–we in fact largely agree that both of these broad accounts (as well as several others) are likely valid. We believe that future research and theoretical synthesis across multiple lines of research is necessary to help determine the degree to which different candidate generalization mechanisms may operate simultaneously, operate on different scales, or be employed under distinct conditions. Here, as the first step, we introduce some of these candidate mechanisms and we discuss the issues currently hindering better synthesis of generalization research. Finally, we introduce some of our own research questions that have arisen over the course of this GAC, that we believe would benefit from future collaborative efforts.

Time separating spatial memories does not influence their integration in humans

“…All this reasoning is centered on the type of integrative encoding mechanisms identified in the related rodent studies [ 3 , 4 ]—in which related experiences are encoded in overlapping sets of neurons—but the performance of our human participants may have been influenced by multiple integrative and interference processes. Indeed, developing clear operational definitions of integration—and generalization more broadly—has been a theoretical thorn in the side of our field [ 65 ] because multiple mechanisms dictate how memories interact at both the time of encoding and retrieval. Broadly construed, behaviours reflecting memory integration could be supported by three encoding processes [ 37 ]: (1) fusing experiences within one representation, potentially at the cost of idiosyncratic experiential details [ 66 , 67 ], (2) explicitly linking two separately represented experiences [ 68 , 69 ], for example, by building a complex situation model [ 70 ], and (3) separately encoding memories that can later be recombined at the time of retrieval [ 71 , 72 ].…”

Reward Uncertainty and Expected Value Enhance Generalization of Episodic Memory

“…We also found that consolidation processing is not necessary for the effect of reward conditioning on episodic memory generalization. We propose that different types of generalization may not be equally affected by consolidation during sleep [ 31 ]. Precisely, the similarity between the conditioned and the generalization stimuli may determine whether the reward effect on generalization is consolidation-dependent.…”

“…If the conditioned and the generalization stimuli are represented in the brain separately (as in an acquired equivalence task), the “offline” state at rest or overnight sleep after conditioning is helpful to integrate or bridge learning experiences together for generalizing or inferencing [ 32 ]. In addition, earlier studies have also proposed a crucial role of sleep in generalization [ 2 , 31 , 33 ]. Conversely, when the difference between conditioned and generalization stimuli is minimal, if the original representation of the conditioned stimuli can be recollected, the representation of the generalization stimuli can also be activated; hence, the consolidation process is not necessary.…”

A roadmap to reverse engineering real-world generalization by combining naturalistic paradigms, deep sampling, and predictive computational models

“…), is a critical feature of biological agents, and a yet unattainable goal for artificial ones. As reviewed by Taylor and colleagues [1], notwithstanding the many different disciplines that have attempted to define generalization and describe its behavioral, cognitive, and neural underpinning, adjudicating between competing theoretical frameworks (e.g., memory integration vs. on-the-fly inference) is currently impossible. Building on the Authors' observation that progress will come from cross-laboratory, cross-disciplinary collaborations, we here suggest additional lines of empirical research that could pave the way to pivotal insights, namely: multimodal stimulation, naturalistic paradigms, deep sampling, and computational models.…”

“…Moreover, generalization can operate on representations evoked in absence of any sensory stimulation, by internal, endogenous processes such as self-generated thoughts which only partially overlap with perception [4], thus calling for dedicated experimental paradigms [5]. Ultimately, to fully understand how real-world generalization works we will need to complement the highly controlled paradigms presented by Taylor and colleagues [1] with experimental conditions that are closer to real life's problem-solving situations. According to basic methodology guidelines, the ability of a study to generalize to other contexts depend on its external validity [6].…”