Publications

You can also find my articles on my Google Scholar profile.

Neural Networks Reveal a Cognitive Continuum Toward Human Abstraction

Li Wenjie , Maggie Henderson , Yonatan Bisk , Jessica Cantlon

The Sixth International Conference on the Mathematics Of Neuroscience and AI, 2025

Do neural network models that fail to behave human-like reflect a fundamental divergence from human cognition, or do they mirror earlier developmental or evolutionary stages? We propose that such models may, in fact, offer insights into the origins of human abstraction. We evaluated over 200 pretrained neural networks alongside macaques, Tsimane adults, US adults and children on three visual match-to-sample tasks targeting increasing levels of abstraction: visual-semantic similarity, shape regularity, and relational reasoning. As task demands grow more abstract, just like monkey’s, model decisions increasingly diverge from adult human behavior. However, representational similarity analyses reveal shared internal structure with all human groups, suggesting overlapping abstraction. We further examine how inductive biases from model designs shape alignment with human cognition. While bigger models sometimes have an advantage with geometric and relational reasoning, it could harm alignment with human semantic structure. We also show that diversity of training data and language supervision understandings of regular geometric shapes in Transformers.

15‐Month‐Olds’ Understanding of Imitation in Social and Instrumental Contexts

Shannon Yasuda , Li Wenjie , Deisy Martinez , Brenden Lake , Moira Dillon

Infancy, 2025

From early in development, humans use imitation to express social engagement, to understand social affiliations, and to learn from others. Nevertheless, the social and instrumental goals that drive imitation in everyday and pedagogical contexts are highly intertwined. What cues might infants use to infer that a social goal is driving imitation? Here we use minimal and tightly controlled visual displays to evaluate 15-month-olds’ attribution of social goals to imitation. In particular, we ask whether they see the very same simple, imitative actions shared between two agents as social or nonsocial when those actions occur in the absence or presence of intentional cues such as obstacles, object goals, and efficient, causally effective action. Our results suggest that infants’ attributing social value to imitation only in the absence of such intentional cues may be a signature of humans’ early understanding of imitation. We propose, moreover, that a systematic evaluation of a set of simple scenarios that probe candidate principles of early knowledge about social and instrumental actions and goals is possible and promises to inform our understanding of the foundational knowledge on which human social learning is built, as well as to aid the building of human-like artificial intelligence.

An Infant-Cognition Inspired Machine Benchmark for Identifying Agency, Affiliation, Belief, and Intention

Li Wenjie , Shannon Yasuda , Moira Dillon , Brenden Lake

Proceedings of the Annual Meeting of the Cognitive Science Society, 2024

Human infants have remarkable abilities to reason about the underlying and invisible causes that drive others’ actions. These abilities are at the core of human social cognition throughout life. Artificial Intelligence (AI) systems continue to fall short in achieving this same commonsense social knowledge. Recent benchmarks focusing on social cognition and theory of mind have begun to address the gap between human and machine social intelligence, but they do not fully consider the social reasoning required to understand scenarios with multiple interacting agents. Building on such benchmarks, we present eight new tasks focusing on different early social competencies, as informed by behavioral experiments with infants. We use a self-supervised Transformer model as a baseline test of our new tasks, and in addition, we evaluate this model on a previous social-cognitive benchmark. While our model shows improved performance on the previous benchmark compared with other data-driven models, it performs sub-optimally on our new tasks, revealing the challenge of learning complex social interactions through visual data alone.

Entropy, mutual information, and systematic measures of structured spiking neural networks

Li Wenjie , Yao Li

Journal of Theoretical Biology, 2020

The aim of this paper is to investigate various information-theoretic measures, including entropy, mutual information, and some systematic measures that are based on mutual information, for a class of structured spiking neuronal networks. In order to analyze and compute these information-theoretic measures for large networks, we coarse-grained the data by ignoring the order of spikes that fall into the same small time bin. The resultant coarse-grained entropy mainly captures the information contained in the rhythm produced by a local population of the network. We first show that these information theoretical measures are well-defined and computable by proving stochastic stability and the law of large numbers. Then we use three neuronal network examples, from simple to complex, to investigate these information-theoretic measures. Several analytical and computational results about properties of these information-theoretic measures are given.