Humans excel at solving novel reasoning problems from minimal exposure, guided by inductive biases, assumptions about which entities and relationships matter. Yet the computational form of these biases and their neural implementation remain poorly understood. We introduce a framework that combines Graph Theory and Graph Neural Networks (GNNs) to formalize inductive biases as explicit, manipulable priors over structure and abstraction. Using a human behavioral dataset adapted from the Abstraction and Reasoning Corpus (ARC), we show that differences in graph-based priors can explain individual differences in human solutions. Our method includes an optimization pipeline that searches over graph configurations, varying edge connectivity and node abstraction, and a visualization approach that identifies the computational graph, the subset of nodes and edges most critical to a model's prediction. Systematic ablation reveals how generalization depends on specific prior structures and internal processing, exposing why human like errors emerge from incorrect or incomplete priors. This work provides a principled, interpretable framework for modeling the representational assumptions and computational dynamics underlying generalization, offering new insights into human reasoning and a foundation for more human aligned AI systems.
When confronted with a new challenge in an unfamiliar and puzzling situation, humans can rapidly formulate a hypothesis based on limited interactions and come up with a solution tailored to the specific problem. This reasoning capability consistently appears in a broad spectrum of human endeavors, ranging from mathematics, science, and technology to social, cultural, and political engagements, and remains a hallmark of human intelligence, yet to be rivaled by state-of-the-art AI (Artificial Intelligence). However, coming up with a quick solution does not guarantee a correct response. It is therefore beneficial to study the extent of human reasoning, to not only inspire the development of artificial machines that reason flexibly and quickly like humans, but to also understand how human reasoning may be impaired in certain circumstances or by certain disorders. To address the extent of human reasoning capability and its applicability to artificial intelligence, we must address three questions: How can humans reason so quickly? What representations or algorithms could lead to reasoning success and failure? And how can these representations and algorithms be realized with neural circuits that could inform the development of intelligent machines?
To tackle the question of reasoning speed, one prominent idea suggests that humans acquire in infancy a set of core knowledge concepts that constrains their expectations of the world and drives their response (Spelke & Kinzler, 2007). Another idea suggests that humans through experience alone learn to form concepts to build a concise world model that could guide their decision making (M. Botvinick et al., 2009;M. Botvinick & Weinstein, 2014;Eckstein & Collins, 2020). These two ideas are not mutually exclusive, as it is possible that a combination of innate knowledge and experience-driven learning is required for the full extent of human reasoning (Kemp & Tenenbaum, 2009). This is the essence of inductive bias, or a set of assumptions other than the data that restricts the learning space and influences the chosen hypothesis in novel circumstances (Baxter, 2000). Inductive biases are characterized as priors in probabilistic models of cognition (Griffiths et al., 2010).
Probabilistic models are sensitive to the exact representation of priors and great success has come from employing logical formulas and programs. For instance, Bayesian program learning exhibited one-shot learning in a handwritten letter task (Lake et al., 2015) and solved novel problems by composing lambda calculus expressions (Ellis et al., 2020). In these frameworks, inductive biases are predefined primitives, also referred to as domain-specific languages, that could be used to construct complex concepts and guide reasoning. These recent successes led some researchers to embrace the long-standing idea that people might employ a similarly discrete and compositional language of thought (Fodor, 1976;Piantadosi, 2021;Quilty-Dunn et al., 2023;Rule et al., 2020). There are strong arguments for having explicit discrete representations to study human reasoning, given that human thoughts are considered compositional, stable and precise (Dietrich & Markman, 2003;Fodor & Pylyshyn, 1988). However, the observation that inductive bias can be represented with discrete programs and symbols does not imply that it has the same representation in the brain.
Inductive bias also found its way into artificial neural networks, most often encoded in the network’s architectures to exhibit properties that could constrain the learning process and lead to better generalization (Goyal & Bengio, 2022). For example, convolutional neural networks have translational invariance, meaning they can recognize patterns regardless of their location within the input (Krizhevsky et al., 2017;Lecun et al., 1998). In contrast, graph neural networks (GNNs) encode relational inductive biases, which prioritize the relationships and connections between entities in a dataset (Battaglia et al., 2018). Artificial neural networks with built-in biases have achieved great success in some reasoning tasks (An & Cho, 2020; D. G. T. Barrett et al., 2018;Jahrens & Martinetz, 2020;Kerg et al., 2022;Małkiński & Mańdziuk, 2024;Sinha et al., 2020;Webb et al., 2023). The lack of discrete representations in these neural networks also challenge the language of thought hypothesis. Indeed, the current best models of language itself are deep networks with continuous internal representations (Bubeck et al., 2023). Unfortunately, interpreting the internal representations of neural networks remains a major challenge for the field (Ghorbani et al., 2018), and aligning artificial circuits to the biological circuits employed by humans is no simple feat (D. G. Barrett et al., 2019).
Graph theory might be a potential bridge between discrete representations and neural networks, bringing interpretability of function to connectivity and topology. First and foremost, graphs are go
This content is AI-processed based on open access ArXiv data.
