Examples
Here are some quick tutorials on how to do specific things in
Simbrain. The basics of building a network and other
simple examples are covered in the quick-start.
Plot the activation of a node over time with a time-series plot.
1) Create a neural network and a time series plot.
2) Create a coupling
from a node in the neural network to a time series in the
plot. The easiest way to do this is as follows. Right
click on the node whose activation you want to plot, and in the context
menu go to Send coupling to > TimeSeriesPlot1 > Series0.
(TimeSeriesPlot1 is the name of the time series plot. If you have
multiple plots open it might be a different name. Series0 is a
particular time series in the plot.) By repeating this for
different nodes you can plot the activation of multiple neurons at
once. If you want to add more time series beyond the default 5, click "add" in the time series component.
3) Now run your simulation and you should see a plot of the changing activation of the neuron. For some purposes it helps to turn "auto range" off in the time series preferences, and manually set upper and lower bounds.
Plot the activation of a set of nodes using a projection plot and the coupling manager
(There are other ways to do this, some of them simpler in some cases, but this technique also shows you how to use the coupling manager). This assumes a network component is open in the desktop.
1) Open a projection plot. To
do this either press on the plot button in the
desktop toolbar, which opens up a drop down menu with different plot
components, or use the menu Insert > New Plot >...
Then select projection plot. The will add a projection plot to
your desktop with a default name like "Projection 1".
2) Open the coupling manager using the Couplings > Open Coupling Manager button.
3) On the left panel of the coupling manager, use the top drop down box
(which allows you to select components) to select the network whose
activity you want to plot.
4) On right panel of the coupling manager, select the plot component you added to the desktop in step 1, e.g. "Projection 1".
5) In the left panel, highlight rows corresponding the neurons whose
activity you are
interested in plotting (do this by clicking on rows while holding the
shift key down). Neurons are labelled by ids. To
find the neurons' ids in the network window you can hover over them and
look at the tooltip.
6) In the right panel of the coupling manager, highlight all the
"dimension" attributes either using the mouse or by pressing command-a
while that panel is in focus.
7) Click "add couplings" at the bottom of the coupling manager window.
Now when you run the workspace each new network state should produce a point in the projection plot.
Note that if you plot more than 25 neurons you will have to add more
dimensions to the projection plot using the add dimension button in the
projection plot, or the menu Edit > Set dimensions... If you use less than 25 neurons you don't have to reset the projection plot but you can if you'd like.
Train a backprop network on a pattern association task
1) Create a backprop network, using the menu Insert > Insert Network > Backprop
2) Enter a topology (a layout for the network) in the dialog that
appears. For example, enter "4,5,4" for a backprop network
with 4 input nodes, 5 hidden layer
nodes, and 4 output nodes.
3) Now you have to create some a
training set: set
of input vectors and a set of target vectors which you want the network
to associate. To do either double click on the backprop tab or right-click on it
and select "Edit / Train Backprop...".
4) In the input data and target data tabs,
edit the cells of the tables to create a pattern assocation task.
Row 1 of input will be associated with row 1 of target, row 2 with row
2, etc.
5) Now all that's left to do is to
actually
train the network. To do this, go to the train tab and
press the run button. This runs the algorithm, which
adjusts
the weights to try to achieve the desired input output
mapping. As the trainer runs, the error should go
down. Once the error gets to an acceptable level (often something
below .1), press the stop button. If you have trouble getting a
low
value you can press the randomize button and try again. Note: Depending
on the associate task
you set up in step 4 you may not be able to achieve a sufficiently low
error value.
7) Now you can test your network to see
how well it did. To do this go to the test data tab, and click
the test row button. This will send each row of data to the
network. The target data you trained it on should appear
with each click of the test row button.