During Stage 1 training, the system learns from labeled inputs with some features specified. In this example, Stage 1 is like a medical school class, where students are taught how to diagnosis patients as normal, hypothermic, or septic. Given both temperature (horizontal axis) and shock (vertical axis) features, the three diagnosis classes are easily separable, but during Stage 1 training, only temperature is specified. Using only temperature, a system can at best achieve a 65% testing accuracy.
Start the movie to see self-supervised ARTMAP learn from Stage 1 training. ARTMAP's long-term memory traces are represented as black category boxes, and the system's classification predictions are shown in red, green, and blue for septic, normal, and hypothermic, respectively. Because shock is never specified during Stage 1, the category boxes are inverted in the vertical dimension, shown as dashed lines.
By the end of Stage 1 training on 50 points, self-supervised ARTMAP has recruited 8 nodes and has a testing accuracy of 65% percent.
During Stage 2 training, the system learns from unlabeled inputs with all features specified. In this example, Stage 2 is like a hospital residency, where new doctors apply their medical school training to diagnosis patients and learn from their own experience. During Stage 2 training, both temperature and shock features are specified, but no diagnosis labels are provided. The system must use its imperfect knowledge about temperature from Stage 1 training to guide learning about shock from these unlabeled inputs and increase classification accuracy from 65% to 100%.
Start the movie to see self-supervised ARTMAP learn from Stage 2 training. Self-supervised ARTMAP uses slow, distributed learning, focused on confident predictions and new features, to incorporate new information about shock into prior knowledge about temperature. Hypothermic and septic category boxes contract toward the top of the plot (high shock), while normal category boxes contract toward the bottom (low shock), slowly increasing testing accuracy. After learning on 1,500 Stage 2 training points, self-supervised ARTMAP has 100% testing accuracy.
Full details about the system, parameters, and training methods are all described in the tech report. These simulations were generated in MATLAB using this script, which requires the self-supervised ARTMAP code.
© 2009 Trustees of Boston University