John A. Bullinaria, Joseph P. Levy

To help understand how semantic information is represented in the human brain, a number of previous studies have explored how a linear mapping from corpus derived semantic representations to corresponding patterns of fMRI brain activations can be learned. They have demonstrated that such a mapping for concrete nouns is able to predict brain activations with accuracy levels significantly above chance, but the more recent elaborations have achieved relatively little performance improvement over the original study. In fact, the absolute accuracies of all these models are still currently rather limited, and it is not clear which aspects of the approach need improving in order to achieve performance levels that might lead to better accounts of human capabilities. This paper presents a systematic series of computational experiments designed to identify the limiting factors of the approach. Two distinct series of artificial brain activation vectors with varying levels of noise are introduced to characterize how the brain activation data restricts performance, and improved corpus based semantic vectors are developed to determine how the word set and model inputs affect the results. These experiments lead to the conclusion that the current state-of-the-art input semantic representations are already operating nearly perfectly (at least for non-ambiguous concrete nouns), and that it is primarily the quality of the fMRI data that is limiting what can be achieved with this approach. The results allow the study to end with empirically informed suggestions about the best directions for future research in this area.