This is a brief introduction about working with the jsWalnut minisimulators here. Each is preconfigured with certain rules and patterns. If you are interested to dig deeper, you can download the walnut.js library and write your own simulations.
To the left in the panels, there are one or more Layers with artificial neurons or nodes. A little menu on top of each layer allows you to choose what parameter of a node is displayed, e.g., Act (for activation), Net (for net input, which is the weighed input activation), etc. Only when Act is selected, you can change the activation of a node by clicking on it. This typically sets it to 1. Clicking it again sets it to -1 (or 0 if that is the lowest allowed activation value). In this way, you can quickly enter simple patterns. A pattern can be saved as described below to create small pattern sets. Once a pattern has been created (or loaded from a pattern set, see below), it can be distorted by pressing the button. Pressing will create a random pattern (set the fraction of activated nodes with the Random Fraction option). Pressing will set all activations to 0 or the lowest value (e.g., -1).
Some panels have ready-made pattern sets, which you can activate by pressing the button in the Pattern Sets panel, e.g.,. This will create two patterns that you can then use. You also create your own patterns by activating a some nodes (artificial neurons) and pressing. This will add a button in the Patterns panel. Then, in the Pattern Menu panel you can set the radio button selection to either Present, Store, Delete, or Change Name. It is currently not possible to save these patterns permanently! In a future version, permanent will be implemented, but right now this is not possible. Pressing the pattern button will execute the selected command for the pressed button. 'Present' will copy the pattern into the neural activations, whereas 'Store' will modify the pattern according to the current activations. It is possible to train the network with these patterns by pressing (i.e., all patterns in the set) or (learn the currentlty activated configuration of activations; e.g., useful with Hebbian type learning). It is possible to for re-use in JavaScript programs. For interactive manipulation this option is not very useful.
The Info panel shows information about the selected Layer, typically its index and name. Then follows information about the currently selected node, e.g., index, activation, and net input. Node parameter information is also available in a tooltip for each node (pause the cursor above a node). A final information category gives general measures of the entire neural network, like (Hopfield) Energy, or the total Error measured over all patterns in the pattern set.
By checking Connection View in the Settings panel, the Layer views change to Weight view. Now, when you click on a node, you see its receptive field, that is the values of all weights on all connections that are connected to the selected node. Weight strength is shown with colors and with a tooltip on each node. The colors may mean different things in different model panels. By selecting Show Values, the values are shown as numbers. This option also works when viewing activation values etc.
You can update the activations in the network by pressing the, where n can be set in the Settings panel (i.e., set Cycle Time). The meaning of this button depends on the neural network paradigm. E.g., in Hopfield networks, which use asynchronous or one-node-at-a-time type updating, a cycle time of 100 means that 100 randomly selected nodes will be updated. But with the perceptron, a cycle time of 1 means that all nodes in the output layer are updated. The button causes a single learning step for the currently activated activations. But the button, runs through all patterns, applies each in turn and learns it; and repeats this m times. Weights can be reset by pressing. It is also possible to remove a fraction of the weights by pressing, where the Lesion Fraction f can be set. If f is 0.80, that means that 80% of all weights are removed. This can be repeated with cumulative effect. For instance, pressing twice will remove about 80% of the remaining 20% of all weights, leaving about 4% of the weights. You can also achieve this by setting the Lesion Fraction to 0.96 and pressing the Lesion button once.