Biomedical Engineering Reference
In-Depth Information
A)
B)
0.05
0.045
0.04
0.035
0.03
LNP Model
0.025
0.02
Linear
Filter
Poisson
Spiking
0.015
Nonlinearity
0.01
0.005
0
0
0.05
0.1
0.15
ISI (s)
FIgURE 2.5:
(a) Poisson distribution of interspike interval for an animal performing a BMI task. (b) LNP
model [
38
].
how modern signal processing can help neuroscience data analysis methods. Indeed, it is our opin-
ion that advanced statistical methods can enormously decrease the amount of data being collected
in the wet laboratory simply by utilizing optimal statistical procedures, which are more efficient
than heuristic approaches.
2.6 NEURal CodINg aNd dECodINg
In the previous sections, we described how information is represented in connectionist networks.
For BMI motor control applications, we will now discuss how representations in populations of
neurons can encode and decode stimuli. The first step is to define what are the basic constructs of
neural coding? From a practical point of view, a neural code must be able to support a sufficiently
rich computational repertoire. It must encode a sufficient range of values with sufficient accuracy
and discriminability. Biological limitations require that the coding scheme must run under a suf-
ficient set of operations implemented with neural mechanics. From the study of cytoarchitecture
and connectionism, the representation of motor commands and sensory stimuli may be distributed
or localized. The landscape of coding depends on how many components are representing the in-
formation, and three classes arise:
•
Local representation: One node is active when a stimulus is presented (cardinal or grand-
mother cells [
39
]). The number of stimuli encoded increases linearly with the number of
cells.
Fully distributed representation: The stimulus is encoded by the combination of all the
neurons.
Sparsely distributed representation: Only a fraction of the neurons encode the stimuli.
•
•
