Extreme Point Machine (EPM): 2-D 2-Class Version
by Ahmad Masadeh and Mohamad Hassoun (May 1998)
This (restricted complexity) version of our EPM employs a polynomial preprocessing (mapping) layer or a mapping layer of perceptrons with random Gaussian weights. Training the output perceptron
weights is accomplished using the AHK learning rule (click here
1. Use the mouse to create a Training Set: point and click (Class 1 = red square, Class 2 = black square). Please note that
the coordinates of the input window are [-5,+5] for both axis. The training set is restricted to a maximum of 150 points. The training set must contain at least one sample per class.
An interesting two interwind spirals (of period 2) training set may be automatically loaded by clicking on the "Load Spirals" button. However, the user is not allowed to use this option until at least one user specified training set (a simple one is recommended) is entered and trained (or cleared).
2. If you make a mistake, click on Clear. This will clear the training set, reinitialize margins, and initialize weights to zero. You may also selectively delete a training sample by clicking on it once (click the center of the square representing such sample).
3. Set Polynomial Order (r). You may use an integer between 1 (linear) and 10. If you choose to use a mapping layer of perceptrons, then set the standard deviation s of the random Gaussian weights (where s is a value between 1 and 10), the number of hidden units J (min 2 and max 150), and your choice of hidden unit activation function (sign quantizer or hyperbolic tangent).
4. Press the Train button to start training. The initial training phase assumes zero initial weights and it defaults to minimum of 2000 cycles and maximum of 3000 cycles. The maximum number of cycles can be set (an integer between 1 and 50,000) in the "Max Cycles" window.
The cycle number is displayed in the cycles window. Train allows you to continue training strating from the state (weights) where the previous training
session ended. At the end of training the applet highlights (in green) the separated regions. Extreme Points (support vectors) from
each class are shown with a "ring" around them only when a solution is found. Note that these marked points are only meaningful after a sufficient number of training cycles is executed so that the separating surface converges to its terminal configuration. The user can make sure that this state is reached by further training to assure that the separating surface has converged. Misclassified samples from class 1
and class 2 are designated by a "hollow red square" and "hollow black square", respectively.
5. The user may at this point add or delete samples and/or change polynomial order (or the perceptron hidden layer parameters) and continue training if he/she desires. However, doing so invokes automatic reset (i.e., output unit weights are cleared, margins are reset, and cycles counter is reset to zero). This "Auto Reset" also leads to a new set of hidden layer random weights for the case of the EPM with perceptron preprocessing layer.
6. The Reset button reinitializes the weights (and the training cycle count) to zero, resets margins and clears the separating surface, but retains the training set. Reset leads to a new set of hidden layer random weights for the case of the EPM with perceptron preprocessing layer.
7. Execution Time: A Gateway G6 333MHz PC executes around 4000 cycles/min a 150 sample problem with polynomial preprocressing and r=10 (same problem executes at 1600 cycles/min when a 150 unit perceptron preprocessing layer is employed).
The interwind period-2 spirals problem can be solved (with high probability) in 1000-3000 cycles (20 to 60 seconds on a 333MHz PC) employing EPMs of about 100 units perceptron preprocessing layer (with tanh activations and Gaussian random weights having standard deviation between 5 and 10). An order 10 polynomial EPM can solve this spirals problem in about 7000 cycles (equivalent to approximately 90 seconds execution time on a 333MHz PC). It should be noted that the present applet code is not optimized for speed.