Novel VQ Designs for Discrete HMM On-Line Handwritten Whiteboard Note Recognition

Neural net vector quantizers for discrete HMM-based on-line handwritten whiteboard-note recognition

“…In this section we briefly explain vector quantizers (VQs), review the codebook switching as presented in [4], and describe the notations.…”

“…However one might argue that there are more sophisticated VQs than the k-means based. In this paper we therefore evaluate and confirm the findings of [4] by using two neural net based VQs, implementing the "Winner-Take-All" (WTA) update rule and the "Neural Gas" (NG) approach, respectively. One major outcome is that while both neural net VQs perform better than the k-means VQ, their performance can still be improved by the codebook switching approach.…”

“…Thereby it has been observed that the pressure information of the trajectory gets lost due to quantization error. To overcome this effect, in [4] we present a novel VQ scheme using switching codebooks, wherein the pressure information is kept without any loss and the statistical dependencies between the pressure and the remaining features are modeled. However one might argue that there are more sophisticated VQs than the k-means based.…”

“…In our previous work (see [4]), the use of discrete HMMs in on-line handwriting recognition is further investigated using a standard k-means algorithm (see [5]) as VQ. Thereby it has been observed that the pressure information of the trajectory gets lost due to quantization error.…”

“…Section 3 summarizes vector quantization in general and the WTA and NG approaches in particular. Then, codebook switching as presented in [4] is reviewed. In an experimental section, both neural net VQs compete against each other and against a standard kmeans VQ.…”