gives rise to a floating point codebook. The linearly
decaying learning kernel only makes it possible to realize
the algorithm on a hardware platform. This comes from the
fact that it employs only integer operations and bit rotations.
That is to say the form of Eq. 2 is changed to (2n - k)
n, where is the left-shift operator. According to the
experimental results, it is surprising to observe that when it
comes to an integer codebook, the performance of the
linearly decaying learning kernel is superior to both the
non-linear learning ones. In this case, the accumulation of
truncation and rounding errors causes the higher MSEs for
the floating point codebook.