Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks

Paper · arXiv 2402.01821 · Published February 2, 2024
Cognitive Models LatentKnowledge GraphsNatural Language Inference

large language models can generate cognitive tasks, specifically category learning tasks, that match the statistics of real-world tasks, deriving rational agents adapted to these tasks using the framework of meta-learning, leading to a class of models called ecologically rational meta-learned inference (ERMI). ERMI quantitatively explains human data better than seven other cognitive models in two different experiments. It additionally matches human behavior on a qualitative level: (1) it finds the same tasks difficult that humans find difficult, (2) it becomes more reliant on an exemplar-based strategy for assigning categories with learning, and (3) it generalizes to unseen stimuli in a human-like way.

Ecological rationality refers to the idea that humans are rational agents adapted to the ecological environments they interact with.

We show that large language models (LLMs) – having been trained on large amounts of human-generated text – can serve as a useful tool for generating ecologically valid tasks, thereby addressing the first challenge.

We illustrate our approach using the domain of category learning (Ashby & Maddox, 2005) — one of the best-studied areas of cognitive science.

Unlike Bayesian models, meta-learned models of cognition can learn adaptive priors by repeatedly interacting with a distribution of tasks. Furthermore, these models have been shown to converge onto the optimal learning algorithm for the environments they are trained on (Ortega et al., 2019) and can be used in cases where the hand-crafting of assumptions is impractical or even infeasible

To illustrate one example, based on this prompt, the model constructed a category learning task with [SODIUM, FAT, PROTEIN] content as feature names and [HEALTHY, UNHEALTHY] as category labels.

During evaluation – i.e., once training is completed – the neural network implements a free-standing learning algorithm that can predict the category label of a new stimulus based on preceding stimulus-category pairs, despite its parameters being frozen.