Figure 3 |. Weights of decoding and encoding models are difficult to interpret.

Three examples (rows) illustrate the difficulty of interpreting decoder weights for a pair of voxels. In the first example (top row), only the top voxel contains signal (stimulus information, red) and the two voxels have independent noise. This scenario is unproblematic: both univariate mapping (second column from the left) and decoder weight maps detect the informative voxel (red). In the second example (second row), both voxels contain the same signal. Here, univariate mapping and weight maps often work. However, the LASSO decoder, because of its preference for a sparse solution, may choose one of the voxels arbitrarily. In the third example (third row), only the top voxel contains signal and both voxels contain correlated noise. Univariate mapping correctly identifies the informative voxel. Linear decoders will give negative weight (blue) to the uninformative voxel, so as to cancel the noise.