Cascading Machine Learning Model Modulation: a New Approach to Increasing the Width of Applications

 

Machine learning models are often performing fabulously in narrowly defined domains. However, once we apply to model a different domain, the system needs to be re-trained or perhaps fundamentally changed by selecting a better algorithm for the problem at hand.

Ideas that can help the inductive learning model be applicable in wider domains have been discussed in major publications. They usually rely on creating ensemble systems containing of multiple, mutually-supporting modules that improve the predictive accuracy.

An idea that I’m working on is somewhat novel. It borrows from the frequency modulation field (FM) applied in such areas as circuit design, telecommunication, or electronic musical instrument production. The concept is based on the fact that a function representing a performance of an oscillator can be radically modified by applying multiple modulation functions. In other words, the base model (a.k.a. the carrier wave) is subjected to a series of modifications that can be stacked up to accomplish a new objective. This procedure has many parallels in machine learning, where modeling a data set can be viewed as developing a hyperplane separating categories of the response variable. The process of stacking up model modulations conceptually resembles hidden layer functionality in neural networks. This is not the first time when linear algebra helps to generate more model complexity. However, the model modulation concept does not extend only to neural architectures. This is where the musical examples are particularly inspiring. We can think of machine learning models as complex functions subjected to cascading external modulation rather than internal modifications. If we can achieve a reasonable modulation level that extends the applicability of the “carrier” model, then a significant improvement of the general model coverage can be accomplished.

Machine learning model modulation is one of the more exciting concepts that I worked on for a while. I hope to be able to present some results soon.  

#ai #machinelearning #FM #frequencymodulation #datamodeling #datamining