How much information is associated with a particular stimulus?

Although the Shannon mutual information can be used to reveal general features of the neural code, it cannot directly address which symbols of the code are significant. Further insight can be gained by using information measures that are specific to particular stimuli or responses. The specific information is a previously proposed measure of the amount of information associated with a particular response; however, as I show, it does not properly characterize the amount of information associated with particular stimuli. Instead, I propose a new measure: the stimulus-specific information (SSI), defined to be the average specific information of responses given the presence of a particular stimulus. Like other information theoretic measures, the SSI does not rely on assumptions about the neural code, and is robust to non-linearities of the system. To demonstrate its applicability, the SSI is applied to data from simulated visual neurons, and identifies stimuli consistent with the neuron's linear kernel. While the SSI reveals the essential linearity of the visual neurons, it also successfully identifies the well-encoded stimuli in a modified example where linear analysis techniques fail. Thus, I demonstrate that the SSI is an appropriate measure of the information associated with particular stimuli, and provides a new unbiased method of analysing the significant stimuli of a neural code.