Error Resilient Real AdaBoost

AdaBoost is a binary classification algorithm that combines several weak classifiers into one strong classifier. This algorithm has relatively good results, even for nearly random base classifiers. Ever since it was published, many variants of the algorithm have been developed for different specific cases. In this project, we focus on a specific version of the algorithm, Real AdaBoost, in which the output of each weak classifier is a real number. Each number represents the confidence level of the classifier in the specific classification decision, and the final classification result is the sum of outputs of all the classifiers.

One of the major problems in AdaBoost-based algorithms is the severe performance degradation when operating in noisy environments. In this project, we research two models for a noisy communication channel between the weak classifiers and the central computing unit. The research question is – what is the degradation of the algorithm performance, and how can we improve it under certain assumptions on the noise model?

For AWGN, we minimized the mismatch probability due to noise by calculating an optimal set of weights for the weighted sum of the weak classifiers. By solving an optimization problem, we successfully increased the noise-resilience of the algorithm. We show a significantly improved performance over the noisy channel for three different databases. The second noise model we considered is due to quantization of the outputs of the weak classifiers. In this case, we were not able to outperform the original algorithm, but managed to reach similar performance using only few representation bits per classifier.

classifier

Error Resilient Real AdaBoost