Foundations of Neuronal Representations - Brain-Machine Interface Engineering - page 32

Within each of these representational schemes, several operations are necessary to support a real-
istically complex computation of sensory information and internal representations, including the
ability to combine, match, map, inverse map, compete, attenuate, and scale, to name a few. In gen-
eral though, all aspects of our behavior are ultimately governed by the responses of neurons, which
operate by changing their electrochemical makeup (polarizing and depolarizing). To measure how
strong the relationship is between a stimulus and neural response, one can measure the correlated
electrical activity between the stimulus and response. Some have framed the functional neural rep-
resentation problem from a variety of perspectives, which are summarized in Table 2.1 . The table is
basically broken into two parts (coding and decoding), which depend upon what quantity is to be
determined by the experimenter. In neural coding, a stimulus (visual and proprioceptive) is given
to the system, and the neural response is to be determined. The decoding problem is the opposite,
where a neural response (i.e., motor intent) is provided by the subject, and the experimenter has to
determine the physical representation of the neural modulation. There are many experimental and
computational approaches to elucidating the neural code that are not mutually exclusive!
The protocol that one uses to deduce the relationships in the input or output neuronal data
generally falls into two classes: exploratory analysis and inference of electrophysiological recordings.
Exploratory analysis is highly descriptive where graphs and numerical summaries are used to describe
the overall pattern in the data. For motor BMIs, some examples of the goals of the exploratory analy-
sis could include characterizing the spiking and LFP correlation for different behavioral conditions or
explaining the spiking variation for certain behaviors in a data set. Exploratory analysis often involves
characterization of a statistical distribution and can include computations such as spike rates, evoked
LFP, spike-triggered averages, LFP spectral analysis, and spike-field coherence. Exploratory analysis
often seeks to find patterns in nature, which can be inherently subjective. In contrast, inference is
strongly hypothesis driven, where the goal is to produce neural and behavioral data in a specific way
on a small scale and then draw more general conclusions about motor control for BMIs. The analysis
often involves finding functional relationships in neural recordings with external quantities. Tuning
curves (described later) have been used to explain the preferred kinematic spike rate for a neuron

TaBlE 2.1: Intersection of experimental variables for
neural coding and decoding

Stimulus Neural Response

Coding Given To be determined

Decoding To be determined Given