Tom Griffiths

We use language to communicate our thoughts. But is language merely the expression of thoughts, which are themselves produced by other, nonlinguistic parts of our minds? Or does language play a more transformative role in human cognition, allowing us to have thoughts that we otherwise could (or would) not have? Recent developments in artificial intelligence and cognitive science have reinvigorated this old question. Could language hold the key to the emergence of both artificial intelligence and important aspects of human intelligence? The four contributions in this symposium address this question by drawing on behavioral and neural evidence from people, and the remarkable recent developments in AI which appear to show that artificial neural networks trained on language come to have an astonishing range of abilities. Despite the diversity of the speakers' perspectives, the four contributions paint a coherent (if complex) picture: The abilities of large language models (LLMs) serve as an existence proof of just how much can in principle be learned from language. LLMs also act as a stress test of cognitive theories. The evidence of neural dissociation between linguistic and conceptual processing points to the multiple realizability of human-like cognition. Finally, there is an acknowledged need for systematic research on how the successes and failures of LLMs inform our understanding of human cognition.

Human planning tends to be efficient, focusing on a relatively small number of options when considering future paths. Recent proposals have suggested that this efficiency reflects intelligent deployment of the limited resources available for planning. A prediction of this and related proposals is that when individuals spend time thinking should depend on the benefits and costs of additional computation. We tested this hypothesis by measuring how much time humans spent thinking before acting in over 12 million online chess games. Players spent more time thinking in board positions where additional computation was more beneficial. This relationship was greater in stronger players, and was strengthened by considering only the information available to the player at the time of choice. A simple model based on measuring the actual cost of spending time thinking in online chess was able to capture qualitative features of this relationship. These results provide evidence that the amount of time humans spend thinking is appropriately sensitive to the value of computation.

Humans organize semantic knowledge into complex networks that encode relations between concepts. The structure of those networks has broad implications for human cognitive processes, and for theories of semantic development. Evidence from large lexical networks such as those derived from word associations suggest that semantic networks are characterized by high sparsity and clustering while maintaining short average paths between concepts, a phenomenon known as a “small‐world” network. It has also been argued that those networks are “scale‐free,” meaning that the number of connections (or degree) between concepts follows a power‐law distribution, whereby most concepts have few connections, while a few have many. However, the scale‐free property is still debated, and the extent to which the lexical evidence reflects the naturally occurring semantic regularities of the environment has not been investigated systematically. To address this, we collected and analyzed semantic descriptors, human evaluations, and similarity judgments from four large datasets of naturalistic stimuli across three modalities (visual, auditory, and audio‐visual) comprising 7916 stimuli and 610,841 human responses. By connecting concepts that co‐occur as descriptors of the same stimuli, we construct “multimodal” semantic networks. We show that these networks exhibit a clear small‐world structure with a degree distribution that is best captured by a truncated power law (i.e., the most‐connected concepts are less common than predicted by a perfect power law). We further show that these networks are predictive of human sensory judgments on these domains, as well as reaction times in an independent lexical decision task. Finally, we show that multimodal networks also share overlapping themes with previously analyzed lexical networks, which upon a more rigorous reanalysis are revealed to be truncated too. Our findings shed new light on the origins of the structure of semantic networks by tying it to the semantic regularities of the environment.

A hallmark of effective teaching is that it grants learners not just a collection of facts about the world, but also a toolkit of abstractions that can be applied to solve new problems. How do humans teach abstractions from examples? Here, we applied Bayesian models of pedagogy to a necklace-building task where teachers create necklaces to teach a learner “motifs” that can be flexibly recombined to create new necklaces. In Experiment 1 (N = 151), we find that human teachers produce necklaces that are simpler (i.e., have lower algorithmic complexity) than would be expected by chance, as indexed by a model that samples uniformly from all necklaces that contain the target motifs. This tendency to select simpler examples is captured by a pedagogical sampling model that tries to maximize the learner's belief in the true motifs by prioritizing examples that have fewer alternative interpretations. In Experiment 2 (N = 295), we find that simplicity is beneficial. Human learners recover the underlying motifs better when teachers produce simpler sequences, as predicted by the pedagogical sampling model. However, humans learn best from human teachers rather than from model-generated examples, which suggests that human teachers have additional expectations about how learners will interpret examples that are not captured by standard models of teaching. Our work provides a principled framework to understand when and why teachers use simple examples to convey abstract knowledge.

Inferring an individual's preferences from their observable behavior is a key step in the development of assistive decision-making technology. Although machine learning models such as neural networks could in principle be deployed toward this inference, a large amount of data is required to train such models. Here, we present an approach in which a cognitive model generates simulated data to augment limited human data. Using these data, we train a neural network to invert the model, making it possible to infer preferences from behavior. We show how this approach can be used to infer the value that people assign to food items from their eye movements when choosing between those items. We demonstrate first that neural networks can infer the latent preferences used by the model to generate simulated fixations, and second that simulated data can be beneficial in pretraining a network for predicting human-reported preferences from real fixations. Compared to inferring preferences from choice alone, this approach confers a slight improvement in predicting preferences and also allows prediction to take place prior to the choice being made. Overall, our results suggest that using a combination of neural networks and model-simulated training data is a promising approach for developing technology that infers human preferences.

How do cognitive pressures shape the lexicons of natural languages? Here, we reframe George Kingsley Zipf's proposed “law of abbreviation” within a more general framework that relates it to cognitive pressures that affect speakers and listeners. In this new framework, speakers' drive to reduce effort (Zipf's proposal) is counteracted by the need for low‐frequency words to have word forms that are sufficiently distinctive to allow for accurate recognition by listeners. To support this framework, we replicate and extend recent work using the prevalence of subword phonemic sequences (phonotactic probability) to measure speakers' production effort in place of Zipf's measure of length. Across languages and corpora, phonotactic probability is more strongly correlated with word frequency than word length. We also show this measure of ease of speech production (phonotactic probability) is strongly correlated with a measure of perceptual difficulty that indexes the degree of competition from alternative interpretations in word recognition. This is consistent with the claim that there must be trade‐offs between these two factors, and is inconsistent with a recent proposal that phonotactic probability facilitates both perception and production. To our knowledge, this is the first work to offer an explanation why long, phonotactically improbable word forms remain in the lexicons of natural languages.