Viewpoint on Computational Modelling of Multisensory Integration

This article presents a compiled interview with three prominent figures in multisensory research: Ulrik Beierholm, Hans Colonius, and Ladan Shams. Through structured dialogue, the authors reflect on the evolution of computational modeling, from early mathematical psychology to the current dominance of Bayesian Causal Inference (BCI).The discussion highlights key paradigm shifts, specifically the transition from descriptive studies of illusions to normative, mechanistic explanations of perception. The authors identify the Ernst and Banks (2002) demonstration of statistically optimal integration as a pivotal milestone in establishing how the brain weighs sensory reliability. Crucially, the authors discuss the subsequent development of BCI (Körding et al., 2007) as the first framework to formally address the "binding problem"—the computational challenge of determining whether sensory signals originate from a common source (integration) or distinct sources (segregation).Current trends and controversies are explored, including the application of computational phenotypes to clinical populations, the debate over the neural implementation of Bayesian computations, and the rising influence of deep neural networks. While AI offers new avenues for analyzing high-dimensional data, the authors emphasize the need for biological realism and explainability. The dialogue also addresses the importance of Open Science, the challenge of accounting for individual differences, and the necessity of moving toward naturalistic environments.The article concludes with professional advice for early-career scientists, emphasizing interdisciplinary training in mathematics and statistics. Collectively, the authors argue that multisensory processing is not a "special case" of perception but the default mode of brain operation, providing essential principles for understanding learning, plasticity, and memory.