Ensemble Inference Cores
All ensemble systems consist of two key components. First, a strategy is needed to build an ensemble that is as diverse as possible. Some of the more popular ones, such as bagging, boosting, AdaBoost, stacked generalization, etc.
A second strategy is needed to combine the outputs of individual classifiers that make up the ensemble in such a way that the correct decisions are amplified, and incorrect ones are canceled out. Several choices are available for this purpose as well.
Two interrelated questions need to be answered in designing an ensemble system:
- How will individual classifiers (base classifiers) be generated?
- How will they differ from each other?
The answers ultimately determine the diversity of the classifiers, and hence affect the performance of the overall system. Therefore, any strategy for generating the ensemble members must seek to improve the ensemble’s diversity.
The process of creating an ensemble is a time consuming one, requiring significant computational resources. Therefore, ensembles are typically created using programs running at some powerful PCs or even supercomputers. This means that creating ensemble classifiers in some embedded system using this approach would not be feasible. To meet this need, So-Logic has developed an IP cores that allow ensemble classifiers to be created directly in hardware. These cores offer significant improvement in time required to build an ensemble, compared with traditional software approaches.
So-Logic currently offers the following ensemble building cores:
- Decision Tree Ensemble Inference Core - this core can be used to build an ensemble composed from decision trees directly in hardware