Intuition in Silico: Representational Alignment of Deep Neural Networks with Human Brain Dynamics in Intuitive Reasoning

Abstract

Predictive reasoning—the capacity to anticipate future states—lies at the core of both biological cognition and artificial intelligence, guiding decisions in diverse domains ranging from object manipulation to strategic planning. In humans, a growing body of research points to a specialized “intuitive physics network” in parietal and frontal regions that supports rapid judgments about stability, collisions, and other physical events. However, standard deep neural networks (DNNs) trained on image recognition tasks often fail to replicate such nuanced physical inference; they can succeed at object categorization without internalizing the causal and dynamic structure underlying real-world interactions. This thesis investigates whether tailoring DNN training objectives—via explicit stability classification—can produce internal representations more aligned with the human intuitive physics network. We employ Representational Similarity Analysis (RSA) to compare layer-wise activations from multiple network architectures (including untrained, ImageNet-pretrained, and video-pretrained models) to pre-existing functional Magnetic Resonance Imaging (fMRI) data collected as participants judge block-tower stability. By correlating model-based and brain-based representational dissimilarity matrices, we measure how closely each network’s encoding mirrors neural activity in parietal regions. Preliminary findings suggest that specialized “physics modules,” or other task heads, can enhance alignment with dorsal-stream fMRI signals compared to conventional architectures. This highlights the importance of task-specific learning objectives for capturing human-like physical inference. More broadly, the findings inform debates about whether the human brain’s intuitive physics emerges from a dedicated forward-simulation mechanism or from a hybrid model that blends simulation with learned heuristics.