Using a simple neuron model we examine how different spike patterns such as bursts encode signals. First we used the direct method to estimate the information carried in bursts and spikes. We found that single spikes contain significantly more information per spike than bursts spikes. Furthermore the majority of the information in bursts is carried by the first spike. However, the timing of bursts is much more robust against noise than single spike times. Thus bursts are more efficient (epsilon = I / H) information carriers (Kepecs, Garibay & Wang, unpublished).

However, our comparison using the direct method can be misleading. A quantitative comparison of entropy rates is difficult when the firing rates are not equal. In our model (as well as in published data) single spikes rates are very different from burst spike rates. To overcome the limitations of this method a new approach is explored which we call ''feature-based'' information. This seeks to determine not only how much but also what kind of information is carried by these different spike patterns.

We define feature vectors as the average stimulus that precedes spikes (Metzner et al, 1998). Instead of only computing a single feature vector we determine one using only single spikes (f_s), one using only burst spikes (f_b) and finally one using bursts as units (f_bu, e.g. first burst spikes). We then ask how well do different kinds of spikes signal particular input features? This ``feature-based'' information is a more natural framework for a comparison of spikes with bursts since it eliminates the requirement of equal firing rates. We found that spikes and bursts indeed carry different kinds of information as defined by their distinct feature-vectors. Spikes carry more information about f_s than f_b and vice versa, bursts carry more information about f_b than f_s. We conclude by outlining possible generalizations of this method to multiple features and two dimensional stimuli (Reich et al, 2000).