Quick and accurate information calculations based on linear characterizations of sensory neurons

Daniel A. Butts, Adrien E. Desjardins, and Garrett B. Stanley
Division of Engineering and Applied Sciences
Harvard University

---

Information theory provides assumption-free measures of the encoding properties of neurons, but as a result of remaining assumption free, these measures often require a prohibitive amount of data to properly estimate. We derive a model-based information calculation (MBIC) that uses a neuron's linear kernel and its non-linear mapping to a firing rate, allowing the instantaneous information rate of a neuron to be calculated using a fraction of the data required by existing direct methods. We also use the instantaneous information rate to evaluate how well the input-output relationship of a neuron is captured by models of sensory encoding. We find that a strict application of this quasi-linear model to real neurons does not fully capture the information encoded by these neurons, but that a simple adjustment of the model's non-linear mapping that takes into account the neuronal refractory period can correct this and properly estimate the information. We apply these methods to models of visual neurons, where we calculate information rates for spatiotemporal input. Thus, this technique allows information calculations in a variety of systems where such characterizations are currently impractical, and furthermore explicitly relates easily measurable properties of a neuron's encoding to its ability to transmit information.