​​​​​​​Talk from Martin Hebart – Max Planck Institute, Leipzig​​

Short abstract:​​

A central aim in the cognitive neurosciences is to reveal the nature of the representations in our minds and brains. While much recent research has focused on studying and comparing representations through the lens of representational geometries, in our work we aim to move one step further and uncover interpretable dimensions underlying these representations. To illustrate the value of the approach, I will highlight multiple research directions in our lab focused on the processing of visually perceived objects. I will first show work that has revealed dimensions underlying our mental representation of thousands of objects, reflecting interpretable visual and conceptual object dimensions that capture most explainable variance in similarity judgments at the individual trial level. Building on these dimensions, I will use densely sampled fMRI to demonstrate that traditional notions of object representations in the visual system focused on object category may be a special case of a broader representational space of object dimensions. Finally, I will demonstrate how this approach can teach us about the similarities and differences in human and macaque neural representations, and I will reveal how deep neural networks suffer from a visual bias that is different from a human semantic bias when representing objects. Together, the aim of this talk is to highlight the opportunities offered by moving towards studying the dimensions underlying representations in biological and artificial neural systems.​