ACTIVE LEARNING via SUPPORT VECTOR MACHINES


Active Learning is a machine learning technique where landscape points are "labeled".  Each landscape point corresponds to a set of input parameters, and the label is the resulting output of a function, simulation, or process.  Active Learning attempts to identify a region of the output landscape that meets or exceeds a certain criteria, as efficiently as possible (i.e. by labeling as few points as possible).

The material below was provided by Dominic Mazzoni, of JPL's Machine Learning Group.  We use Active Learning at SwRI as a way to drastically improve the results and efficiency of complex numerical simulations.  The Active Learning concepts, as they pertain to simulations, have been developed, enabled, and supported by an outstanding collaboration with JPL's Machine Learning Group.  We gratefully acknowledge their contributions!

This image represents a function that we are trying to learn:

Think of it as a 2-dimensional analog to the simulator problem; we are trying to learn when the function is white or black as a function of the x, y coordinates - similar to predicting whether the result of a simulation is good or bad based on its input parameters, but easier to visualize because it's only 2-dimensional.

The toy function is trivial to solve in polar coordinates but tricky in cartesian coordinates, which is why it makes a good example.

Movie 1 shows an attempt to learn this function using Active Learning, up through 500 points. The current support vectors are shown in green - note that as new points are learned, old support vectors are sometimes dropped, too...

Movie 2 shows the first 150 points of the same simulation, but this time comparing Active Learning (on the left) with random point selection on the right. The bottom plots show the binary output, and the top plots show confidence, too.

Movie 3 shows what happens if you use too much Active Learning - it spends all of its time focusing on a specific region, rather than exploring the rest of the space. Our technique needs to find a balance between focused exploration and balanced exploration of the space. (The SVM looks different here because this uses a Gaussian kernel, the others used a Polynomaial kernel, but otherwise the approach is the same.)

Brian Enke  benke@boulder.swri.edu

Last updated 08-May-2002