This is a visualization of a network I am working on to control a robotic arm. It takes two analog inputs, which are then converted into one-dimensional population codes ("Gauss One" and "Gauss Two"). These are then connected to a two-dimensional population that becomes a "gain-field".

The output will be a weighted sum of inputs from the gain-field, creating a function approximation. In this case, the network will be trained to output joint torque based on current joint angle and desired delta angle.

The code is written in C, compiled on a Mac using Xcode and OpenCV. I will post the code soon, or email me: brian90254 at gmail dawt com.

This is a pattern classification network, or multi-layer "perceptron". What makes it interesting is that it uses a contrastive-Hebbian learning scheme, so there are no credit-assignment problems or back-propagation calculations.

In the "before" visualization, the network is un-trained and presented with a random looking input. The desired output is "H", so the output is clamped during a brief learning period. The weights are adjusted according to a fixed-threshold BCM rule.

After learning, the ouput matches the teaching signal.

Again, this is an Xcode project using OpenCV. Check back for the code, or email me: brian90254 at gmail dawt com.