Figure 2. Confusion Matrices and calculations of Inclusive and Exclusive Categorical Information.

(A) Confusion matrix obtained at the end of the decoding procedure of the spike pattern responses of an example unit. The color in each bin of the matrix represents the joint probability that the neural responses to the stimulus i (rows: actual stimuli) were classified as belonging to the same stimulus (j predicted stimulus and here on the diagonal j=i) or the other stimuli (other columns in that row; j≠i). The Mutual Information (MI) measures the information content of the neural responses using these joint probabilities. MI_max is the theoretical upper bound of MI and depends on the stimulus set size.

(B) Inclusive-Categorical-uniform matrix obtained after a procedure that sets the joint probability outside the categories to equal and average values, effectively canceling out all information in the original confusion matrix (A) that did not pertain to the semantic classification and the classification of single calls within the categories. The mutual information of this modified matrix is called the Inclusive Categorical Information (ICI) of the unit (see methods). Note that the probabilities outside the block diagonal of this matrix are not zero but taken on very small values corresponding to very dark colors on our display.

(C) Exclusive-Categorical-uniform matrix obtained after a procedure that further removes the information in (B) about the classification of calls within categories by setting the joint probabilities within the block diagonal corresponding to each category to the same average value. The mutual information of that modified matrix is the Exclusive Categorical Information (ECI) of the unit. ECI_max is the theoretical upper bound of ECI and depends on the stimulus set size and composition.