Inductive reasoning in minds and machines

Abstract

Induction --the ability to generalize from existing knowledge-- is the cornerstone of intelligence. Large language models have been shown to be capable of certain types of reasoning, however, they are limited in their ability to mimic human induction. In this paper, we combine representations obtained from large language models with theories of human inductive reasoning developed by cognitive psychologists. Our approach can capture several benchmark empirical findings on human induction, and generate human-like responses to natural language arguments with thousands of common concepts and properties. These findings shed light on the cognitive mechanisms at play in human induction, and show how existing theories in psychology and cognitive science can be integrated with leading methods in artificial intelligence, to successfully model high-level cognition.

Using artificial intelligence to fit, compare, evaluate, and discover computational models of decision behavior

“…Some have been using machine learning to discover or approximate the shape of functions that are relevant to decisionmaking, such as utility, probability weighting Peterson et al (2021), and temporal discounting functions (Cavagnaro et al, 2016). Others are using natural language models like vector-based approaches or transformers (Bhatia & Mullett, 2018;Bhatia, 2023) to directly make predictions about human behavior. And yet others are discovering ways to decode neural representations during decision-making (Horikawa et al, 2013;Schönauer et al, 2017).…”

From Natural Language Models to Cognitive Model Validation: A Theoretical Framework

“…The review proposes that NLP models can help to generate experimental stimuli, enabling researchers to acquire naturalistic stimuli while retaining the necessary level of experimental control during lab-based tasks. These stimuli, in turn, can be utilized to construct and inform computational models of cognition (e.g., Bhatia (2022); Bhatia and Richie (2022)) .…”

Verbal Reports as Data Revisited: Using Natural Language Models to Validate Cognitive Models

“…Prior to outlining our framework, it is important to highlight how our approach sets itself apart from the existing literature linking NLPs with human psychology. One focus of this literature is on the similarities or differences between human reasoning and that of NLPs Bhatia (2022); Bhatia and Richie (2022); Bhatia and Stewart (2018); Schulz (2022, 2023). A second focus is to use NLP models to help generate experimental stimuli, enabling researchers to acquire naturalistic stimuli while retaining the necessary level of experimental control during lab-based tasks.…”

“…A second focus is to use NLP models to help generate experimental stimuli, enabling researchers to acquire naturalistic stimuli while retaining the necessary level of experimental control during lab-based tasks. (Stimuli, which, in turn, can be used to construct and inform computational models of cognition (e.g., Bhatia (2022); Bhatia and Richie (2022)) .…”