Digital Signal Processing Reference
In-Depth Information
Figure 8-1.
Cerebellar Model Articulation Controller (CMAC) Block Diagram.
The CMAC network in Figure 8-1 can be viewed as two layers of
neurons, and hence, its operation can be decomposed into two separate
mappings. The input vector is transformed to a vector of binary values,
which, in turn, produces at the output the sum of weights that link itself to
the corresponding input vector of value one. Given an input vector, the
desired output at the output layer can be approximated by modifying its
connection weights through the use of an adaptation process, commonly
known as the training mode in the perceptron theory.
While there have been many studies in the area of CMAC memory and its
associated architectures, most of them were concentrated in adaptive control
problems and a simple two-dimensional CMAC configuration was sufficient
for them.
In the case of signal processing applications, in particular, in speech and
image processing problems because of the data sizes involved and the
locally-stationary and locally ergodic nature of the underlying processes,
there are a number of factors affecting the performance. These, in turn,
require more sophisticated setups. In an earlier work, we have used the
CMAC concept in speech enhancement and echo cancellation with
encouraging results [6]. In particular, the CMAC was used in the Wiener
filtering stage of an Amplitude Spectral Estimation (ASE) for noise and/or
echo cancellation in moving vehicles.
