The above description of the basic algorithm
implicitly introduced all the basic concepts we need to know to
understand how Learning-Tool works:
Input Distribution
The distribution from which the training
(and test) inputs are drawn. In our example the input distributio is
defined byt the following simple procedure (for fixed templates 0
and 1):
Randomly choose template 0 or 1
Add noise (jitter) to each spike in the template
Neural Microcircuit
The circuit which receives the input and
whos response is recorded and analysed (in our example this
a network of 135 leaky-integrate-and-fire neurons).
Response of the Microcircuit
The response (output) of the
neural microcircuit (in our example the 135 spike trains produced by
the microcircuit model).
State of the Microcircuit
The transformed (smoothed) response
(output) of the neural microcircuit (in this examples this
corresponds to a low-pass filtered (30ms) version of the spike
trains). This transformation can also be dropped if one can cope
directly with the spike response.
Sample Time Points
Since we can only handle finite sets of
training examples we must define time points at which we want to
sample the state of the microcircuit (in this example we will sample
the states every 25ms).
Readout Function
A parameterized function/device which gets as
input the circuit states (or in some cases directly the circuit
response) an computes the outputs of the system (n this example a
threshold gate).
Target Function/Filter
A function which defines for each input
time series the target output time series of a readout function. In
mathematical terms this should be a target filter since we are
talking about computations on time series.
Supervised Learning Algorithm
By means of such algorithm the
paramters of the readout -- and only the readout -- are adjusted
such that the actual output of the readout matches as close as
possible the target output.
Training Set
Set of inputs used to determine the parameter of
the readout.
Test Set
Set of inputs different to the training set which is
used to asses the performance of the trained readout.
As we will see each of this terms has its corresponding element within
Learning-Tool .
