What Makes Different People's Representations Alike: Neural Similarity Space Solves the Problem of Across-subject fMRI Decoding

摘要

A central goal in neuroscience is to interpret neural activationnand, moreover, to do so in a way that captures universal principlesnby generalizing across individuals. Recent research in multivoxelnpattern-based fMRI analysis has led to considerablensuccess at decoding within individual subjects. However, thengoal of being able to decode across subjects is still challenging:nIt has remained unclear what population-level regularities ofnneural representation there might be. Here, we present a novelnand highly accurate solution to this problem, which decodesnacross subjects between eight different stimulus conditions.nThe key to finding this solution was questioning the seeminglynobvious idea that neural decoding should work directly on neuralnactivation patterns. On the contrary, to decode across subjects,nit is beneficial to abstract away from subject-specificnpatterns of neural activity and, instead, to operate on the similaritynrelations between those patterns: Our new approach performsndecoding purely within similarity space. These resultsndemonstrate a hitherto unknown population-level regularity innneural representation and also reveal a striking convergencenbetween our empirical findings in fMRI and discussions in thenphilosophy of mind addressing the problem of conceptual similaritynacross neural diversity