On spike threshold and the neural code

Date of Completion

January 2006

Keywords

Engineering, Biomedical|Engineering, Electronics and Electrical

Degree

Ph.D.

Abstract

The information communicated by stereotypical action potentials is thought to be embedded in the timing of the spikes (time code) or their average count (rate code). Here we report a continuum in these encoding mechanisms and by extension, in communicated information between neurons. We demonstrate that auditory midbrain neurons exhibit a tradeoff between spike rate and spike information spanning through each extreme. Feature selective neurons with scarce but temporally precise responses convey higher spike information (bits/spike) than more active neurons with relatively wide receptive fields. Conversely, higher spike rates affect an increase information throughput (bits/sec). We propose and demonstrate with a model that spike threshold accounts for this tradeoff. ^ We then performed in vitro whole cell recordings in pyramidal cells in response to a simulated synaptic input to study the role of threshold in spike timing accuracy. Despite a high degree of covariation between spike threshold voltages and the pre-spike membrane potential history, spiking is reliable and precise. We show that threshold adaptation in a Leaky Integrate Fire model dramatically increases its predictive accuracy. Unusually high average mutual information of 6.60 bits/spike and efficiency of 73%, roughly 2.5 times that observed in viva is also reported. These results suggest that cortical and hippocampal pyramidal neurons are capable of transmitting information at much higher rates than previously thought, and that adaptation in the spiking mechanism enhances the efficiency of the neuronal representation. ^ Lastly we examine the structure of neuronal variability and averaging methods typically used to isolate this variability. While averaging presents a quick and convenient method to cancel the "noise" in repeated neural recordings, we must be mindful that averaging presumes neuronal responses to be Poisson point processes where the average firing rate is the meaningful output metric. We show that the assumption of independence of the neuronal noise from the neuronal response does not apply. We demonstrate that noise is not additive and propose that the threshold nonlinearity of neurons give noise a multiplicative (reliability) and convolution (timing jitter) structure. A novel correlation-based algorithm for extracting jitter and reliability errors is developed and validated by its application to a variety of neural recordings. ^

Link to Full Text

Share

COinS