Fig. 7.

Inverted encoding model. A. Inverting the forward encoding model to reconstruct the stimulus feature, orientation. First, the experimenter specifies some nonlinear transformation of the stimulus into a representational space. Here, orientations of Gabor gratings are transformed into activations on a set of orientation channels C that tile the stimulus space. Then, the fMRI responses B are predicted by solving the linear equation B = WC. To reconstruct stimulus features with a new set of data B₂, we simply invert W to predict a new C₂. B. IEMs allow experimenters to test detailed hypotheses about stimulus representations. Scolari et al. [77] tested the off-channel gain hypothesis (figure adapted with permission). According to this hypothesis, when discriminating between very similar features, it is optimal to enhance the responses of channels close to the relevant feature, rather than directly enhancing the relevant feature. Using an IEM for stimulus orientation, Scolari et al. [77] demonstrated off-channel gain enhancement when subjects performed a difficult orientation discrimination task, compared to when subjects performed a contrast discrimination task on the same stimuli.