A mathematical theory of relational generalization in transitive inference.

Humans and animals routinely infer relations between different items or events and generalize these relations to novel combinations of items. This allows them to respond appropriately to radically novel circumstances and is fundamental to advanced cognition. However, how learning systems (including the brain) can implement the necessary inductive biases has been unclear. We investigated transitive inference (TI), a classic relational task paradigm in which subjects must learn a relation ([Formula: see text] and [Formula: see text]) and generalize it to new combinations of items ([Formula: see text]). Through mathematical analysis, we found that a broad range of biologically relevant learning models (e.g. gradient flow or ridge regression) perform TI successfully and recapitulate signature behavioral patterns long observed in living subjects. First, we found that models with item-wise additive representations automatically encode transitive relations. Second, for more general representations, a single scalar "conjunctivity factor" determines model behavior on TI and, further, the principle of norm minimization (a standard statistical inductive bias) enables models with fixed, partly conjunctive representations to generalize transitively. Finally, neural networks in the "rich regime," which enables representation learning and improves generalization on many tasks, unexpectedly show poor generalization and anomalous behavior on TI. We find that such networks implement a form of norm minimization (over hidden weights) that yields a local encoding mechanism lacking transitivity. Our findings show how minimal statistical learning principles give rise to a classical relational inductive bias (transitivity), explain empirically observed behaviors, and establish a formal approach to understanding the neural basis of relational abstraction.

Focusing on what matters: Modulation of the human hippocampus by relational attention.

Hippocampal episodic memory is fundamentally relational, comprising links between events and the spatiotemporal contexts in which they occurred. Such relations are also important over shorter timescales, during online perception. For example, how do we assess the relative spatial positions of objects, their temporal order, or the relationship between their features? Here, we investigate the role of the hippocampus in online relational processing by manipulating attention to different kinds of relations. While undergoing fMRI, participants viewed two images in rapid succession on each trial and performed one of three relational tasks, judging the images' relative: spatial positions, temporal onsets, or sizes. Additionally, they sometimes judged whether one image was tilted, irrespective of the other. This served as a baseline item task with no demands on relational processing. The hippocampus showed reliable deactivation when participants attended to relational vs. item information. Attention to temporal relations was associated with the most robust deactivation. One interpretation of such deactivation is that it reflects hippocampal disengagement. If true, there should be reduced information content and noisier activity patterns for the temporal vs. other tasks. Instead, multivariate pattern analysis revealed more stable hippocampal representations in the temporal task. This increased pattern similarity was not simply a reflection of lower univariate activity. Thus, the hippocampus differentiates between relational and item processing even during online perception, and its representations of temporal relations are particularly robust. These findings suggest that the relational computations of the hippocampus extend beyond long-term memory, enabling rapid extraction of relational information in perception.

Flexible weighting of diverse inputs makes hippocampal function malleable.

Classic theories of hippocampal function have emphasized its role as a dedicated memory system, but recent research has shown that it contributes broadly to many aspects of cognition, including attention and perception. We propose that the reason the hippocampus plays such a broad role in cognition is that its function is particularly malleable. We argue that this malleability arises because the hippocampus receives diverse anatomical inputs and these inputs are flexibly weighted based on behavioral goals. We discuss examples of how hippocampal representations can be flexibly weighted, focusing on hippocampal modulation by attention. Finally, we suggest some general neural mechanisms and core hippocampal computations that may enable the hippocampus to support diverse cognitive functions, including attention, perception, and memory. Together, this work suggests that great progress can and has been made in understanding the hippocampus by considering how the domain-general computations it performs allow it to dynamically contribute to many different behaviors.

Representation and Computation in Cognitive Models.

One of the central issues in cognitive science is the nature of human representations. We argue that symbolic representations are essential for capturing human cognitive capabilities. We start by examining some common misconceptions found in discussions of representations and models. Next we examine evidence that symbolic representations are essential for capturing human cognitive capabilities, drawing on the analogy literature. Then we examine fundamental limitations of feature vectors and other distributed representations that, despite their recent successes on various practical problems, suggest that they are insufficient to capture many aspects of human cognition. After that, we describe the implications for cognitive architecture of our view that analogy is central, and we speculate on roles for hybrid approaches. We close with an analogy that might help bridge the gap.

Changes in causal attributions and relationship representations: Are they specific or common mechanisms in the treatment of depression?

BACKGROUND: The goal of the study was to examine two central theory-driven mechanisms of change, causal attributions and relational representations, to account for symptomatic improvement in psychodynamic treatment and supportive clinical management, combined with either pharmacotherapy or placebo, in a randomized control trial (RCT) for depression. METHOD: We used data from an RCT for depression, which reported non-significant differences in outcome among patients (N=149) who received supportive-expressive psychotherapy (SET), clinical management combined with pharmacotherapy (CM+MED), or clinical management with placebo pill (CM+PBO) (Barber et al., 2012). Mechanism and outcome measures were administered at intake, mid-treatment, end of treatment, and at a 4-month follow-up. RESULTS: Improvements in causal attributions and in relational representations were found across treatments. Changes in causal attributions did not predict subsequent symptomatic level when controlling for prior symptomatic level. In contrast, decrease in negative relational representations predicted subsequent symptom reduction across all treatments, and increase in positive relational representations predicted subsequent symptom reduction only in SET. LIMITATIONS: The study is limited by its moderate sample size. Additional studies are needed to examine the same questions using additional treatment orientations, such as cognitive treatments. CONCLUSIONS: Findings demonstrate that changes in negative relational representations may act as a common mechanism of change and precede symptom reduction across psychodynamic therapy and supportive case management combined with either pharmacotherapy or placebo, whereas an increase in positive relational representation may be a mechanism of change specific to psychodynamic therapy.