Hierarchical Multisensory Integration: Theory and Experiments

Barcelona, Spain, June 18-19, 2017

The ability to map sensory inputs to meaningful semantic labels, i.e., to recognize objects, is foundational to cognition, and the human brain excels at object recognition tasks across sensory domains. Examples include perceiving spoken speech, reading written words, even recognizing tactile Braille patterns. In each sensory modality, processing appears to be realized by multi-stage processing hierarchies in which tuning complexity grows gradually from simple features in primary sensory areas to complex representations in higher-level areas that ultimately interface with task-related circuits in prefrontal/premotor cortices.

Crucially, real world stimuli usually do not have sensory signatures in just one modality but activate representations in different sensory domains, and successfully integrating these different hierarchical representations appears to be of key importance for cognition. Prior theoretical work has mostly focused on tackling multisensory integration at isolated processing stages, and the computational functions and benefits of hierarchical multisensory interactions are still unclear. For instance, what characteristics of the input determine at which levels of two linked sensory processing hierarchies cross-sensory integration occurs? Can these connections form through unsupervised learning, just based on temporal coincidence? Which stages are connected: For instance, is there selective audio-visual integration just at a low level of the hierarchy, e.g., to enable letter-by-letter reading, or even earlier levels at the level of primary sensory cortices, with multisensory selectivity in higher hierarchical levels then resulting from feedforward processing within each hierarchy, or are there selective connections at multiple hierarchical levels? What are the computational advantages of different cross-sensory connection schemes? What are roles for “top-down” vs. “lateral” inputs in learning cross-hierarchical connections? What are computationally efficient ways to leverage prior learning from one modality in learning hierarchical representations in a new modality?

The workshop will gather a small group of experts to informally exchange the latest ideas and findings, both experimental and theoretical, in the field of multisensory integration. It will consist of two days packed with talks by invited speakers as well as discussions. There will also be a poster session. Researchers, postdocs and graduate students interested in multisensory integration and hierarchical processing are all invited to apply. Click here (http://eventum.upf.edu/event_detail/8963/detail/pire-workshop_summer-school-2017.html) for more information on the event and to register.

This event is jointly organized by the Center for Brain and Cognition at the Universitat Pompeu Fabra in Barcelona and Georgetown University, with funding from the National Science Foundation and the Spanish Ministry of Economy, Industry and Competitiveness.