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.

Keywords: LPFC; generalization; generative adversarial collaboration (GAC); hippocampus; mPFC; memory integration; separate memories; transfer.

Figure 1

Episodic inference tasks. Letters A, B, C, D, E stand for specific stimuli (e.g., an image on the screen for a human, an odor for a rodent). Green circles represent the correct answers. a. In Acquired Equivalence tasks, participants learn that two probe stimuli are paired with the same choice (A-C and B-C), and also that one of the probe stimuli is paired with another choice (A-D). They are then tested to see if they choose D when presented with B (B-D). b. In human versions of associative inference tasks, participants encode overlapping paired associates (A&B and B&C) via observational learning. They are then tested on the relationship between A and C. Most commonly, subjects are presented with an A probe and asked to select between two stimuli, with the choices being the correct C item and another familiar item. c. In the rodent versions of associative inference tasks, subjects learn through reinforcement learning over many trials to choose B over Y when presented with A, and choose C over Z when presented with B (A-B, B-C). They are then tested to see if they will choose C over Z when presented with A, although the A-C choice was never trained. d. On each trial of transitive inference tasks, participants are shown two images on the screen and asked to choose one. Over many trials through corrective feedback, they learn to choose A over B, B over C, C over D and D over E (A>B, B>C, C>D, and D>E). They are tested to see if they will generalize to assume, for example, B>D.

Figure 2

Traditional categorization tasks. Green circles represent the correct answers. a. In the A/B dot pattern stimuli task, via feedback, participants learn to categorize two different groups of dot pattern stimuli (A and B). Each of these categories contain stimuli that share a central tendency, created as distortions from a prototype pattern by moving individual dots with a controlled amount of jitter. Participants are subsequently tested to see if they can accurately categorize: the stimuli from the learning phase; prototypical stimuli from each category; and new stimuli from each category [114]. b. The A/not-A pattern stimuli task only includes examples from a single category during training. At test, new categorical examples are presented as well as new non-categorical examples that are typically random patterns that do not share a central tendency with stimuli from the category or with each other [9]. c. An example of a categorization task that uses stimuli with well-defined binary-value dimensions. Here, cartoon animals vary based on binary characteristics such as color (e.g., yellow/grey), body shape (squared/circular), and/or head orientation (forward/up). Category membership may rely on a single dimension, two dimensions, or more dimensions [85]. Some versions of this type of categorization task are analogous to the dot pattern stimuli task. In this case, one stimulus may serve as the category A prototype and category A exemplars are constructed as its “distortions”, where some of the category-typical features are exchanged for atypical features (stimuli in this figure were adapted from Bozoki etal., 2006 [117]).

Figure 3

The functional category reward prediction task of Pan et al. (2008; 2014). Green circles represent the correct answers. a. During the initial category learning phase, via trial and error, subjects learn to associate groups of stimuli with one another to form two functional categories. If, after seeing one stimulus from a given category (e.g. A1), the subject next selects two other stimuli from that category (e.g. B1 and C1) then they receive a reward. If they do not select stimuli from the same category (e.g. choose B2 and/or C2 after seeing A1) then they are not rewarded. Although not depicted here, subjects learn the stimuli from each group in several different temporal sequences (e.g. A1-B1-C1, B1-C1-A1, and C1-A1-B1). This is to ensure that symmetry is established between category members. b. After initial categories have been well learned, in a separate learning phase, subjects learn to associate multiple novel stimuli with a stimulus from one of the categories (e.g. B1) and multiple other novel stimuli with a stimulus from the other category (e.g. B2). c. At the beginning of each session of the test phase, participants complete several reward instruction trials. In these trials, participants are presented with a stimulus from one category (e.g. C1) along with high reward, and a stimulus from another category (e.g. C2) with low reward. d. After the reward instruction trials in each session of the test phase, participants are tested to see if they will transfer the reward information learned during the reward instruction trials to other members of the same functional category; This is tested for both category members learned during the “initial learning” phase (e.g. to see if they combine A1-B1, B1-C1, and C1-large reward to begin predicting large reward to A1) and for novel category members learned later in the “additional member learning” phase (e.g. to see if they combine N1-B1, B1-C1, and C1-large reward to begin predicting large reward to N1). Rew. = reward.