Joint compression-classification with quantizer/classifier dimension mismatch

In this paper an algorithm is presented to design encoders that achieve good compression and classification. The goal is to minimize the classification error introduced by quantizing the data using encoders operating on low dimension inputs, which are subsets of the high dimension data used by the classifier for classification. In the encoder design information from the other dimensions of the vector is used to develop efficient encoders which are capable of achieving lower classification error for a given distortion. The design allows a trade-off between distortion and classifications costs providing more flexibility in the overall system design. The algorithm is tested on Gaussian mixture data, which is classified using a classifier which takes as input vectors of quantized values. The proposed technique can trade performance to achieve lower complexity, which is desirable in devices having limited computational resources. For 4 dimensional Gaussian mixture data the misclassification was about 2.2% more than that achieved by using encoders of the same dimension as the classifier, while the encoding complexity was reduced by more than a factor of 2.