Biomedical Engineering Reference
In-Depth Information
FIgURE 2.12:
Information theoretic tuning depth in logarithmic scale for 185 neurons computed
from three kinematic variables respectively. The upper plot is tuning depth computed from position, the
middle plot is for velocity, and the bottom plot is for acceleration.
However, in many applications such as the design of neural interfaces, the timing of large
ensembles of cells (hundreds of cells) must be studied in a principled way. Unfortunately, many of
the aforementioned techniques become extremely time-consuming (as all pairs of electrode combi-
nations must be considered) and alternative pattern recognition [
74
] or clustering [
82
] methods to
quantify multichannel prespecified spike timing relations (patterns) have been proposed.
Most statistical analysis of neuronal spike timing assumes Poisson spike generation where
the neural response only depends upon the short-term past of the input stimulus due to the inde-
pendence assumptions embedded in the Poisson model. To estimate the response of a given neuron
to these stimuli, one needs simply to quantify the likelihood of a spike firing in each region of the
stimulus space, which can be estimated by the ratio of multidimensional histograms of spike trig-
gered stimuli. To accurately estimate the histograms, large amounts of data are required to isolate
the analysis to regions of the space where the neuron shows a large response (many counts to evalu-
ate the histograms) or to tailor the stimuli, parameterized by only a few parameters to produce large
neural responses. To overcome data issues, one can time lock the averaging to the spike, a procedure
known as spike-triggered averaging. Once the spike-triggered average is obtained, the neural re-
sponse can be inferred with a nonlinear input output neural model.
