Figure 1 |. Linear encoding and decoding models.

(A) Encoding and decoding model the relationship between stimuli and responses in opposite directions. An encoding model (top) predicts brain responses as a linear combination of stimulus properties (black circles). A decoding model (bottom) predicts stimulus properties as a linear combination of brain responses. (B) Example of a linear decoding model using a 2-dimensional feature space consisting of two voxels. Voxel 1 contains relevant information about which of two classes (green, blue) the stimulus belongs to. Voxel 2 contains no information about the stimulus class. The two dimensions jointly define the linear discriminatory boundary. Note that the weights assigned to each voxel are defined by the vector w, which is orthogonal to the decision boundary. Because the noise is correlated between the voxels, a linear decoder will assign significant negative weight to voxel 2, using this voxel (which contains only noise) to cancel the noise in the voxel which contains signal. As a result, interpreting the absolute weights of linear decoders requires care and additional analyses.