Decoding feature spaces in primate MT: An evaluation of multivariate pattern analysis methods.

In recent years there have been substantial developments in the application of multivariate pattern analysis methods to neuroscience data, e.g. fMRI, EEG/MEG, and single unit recordings. Many of these developments have come from studies examining the brain’s coding of visual objects, a model system in which the underlying coding principles are poorly understood. In contrast, the coding of visual motion in the brain is relatively well known. We know the fundamental features of motion, i.e. direction and speed, and understand how these features of visual motion are coded in MT, the brain area most often associated with representing motion. In the present study, we leveraged our knowledge of visual motion to study the effectiveness of inferential and exploratory methods for analysing multivariate brain recordings. We analysed single unit activity in multielectrode recordings from area MT of sufentanil-anaesthetised marmoset monkeys. The visual stimuli were moving dot patterns varying in speed and direction (7 speeds x 12 directions), or moving grating patterns varying in spatial frequency (SF), temporal frequency (TF), and direction, (3 SF x 3 TF x 12 directions). For each dataset, we used a linear classifier to compute decoding performance from the recordings for all possible pairwise combinations of stimuli. We first applied representational similarity analysis (RSA; Kriegeskorte, 2008), an inferential model testing approach, and confirmed that speed, direction, SF and TF were represented in population response. We then explored the utility of multidimensional scaling (MDS), an exploratory method of “discovering” important feature dimensions in the neural code. MDS was capable of revealing speed and direction feature dimensions for moving dot patterns; for grating patterns, however, MDS conflated SF and TF, and gave no indication that motion direction was relevant. We then applied hierarchical cluster analysis, another exploratory method that assumes categorical structure and aims to recover the categorical divisions. Cluster analysis conflated the feature dimensions in both dot and grating datasets, and provided limited interpretative value. Our findings show that in a well characterised system (visual motion analysis in area MT), where critical stimulus features have been explicitly manipulated, that exploratory analyses have limited value in “discovering” the underlying coding principles used by the brain. Theory and explicit model testing provide better means to uncover the brain’s neural code.