The Cerebellar Model Arithmetic Computer (CMAC) (Albus 1981) is well known as a good function approximator with local generalization abilities. Depending on the smoothness of the function to be approximated, the resolution as the smallest distinguishable part of the input domain plays a crucial role. If the binary quantizing functions in CMAC are dropped in favor of more general, continuous-valued functions, much better results in function approximation for smooth functions are obtained in shorter training time with less memory consumption. For functions with discontinuities, we obtain a further improvement by adapting the continuous encoding proposed in Eldracher and Geiger (1994) for difficult-to-approximate areas. Based on the already far better function approximation capability on continuous functions with a fixed topologically distributed encoding scheme in CMAC (Eldracher et al. 1994), we present the better results in learning a two-valued function with discontinuity using this adaptive topologically distributed encoding scheme in CMAC.

Predictability minimization (PM—Schmidhuber 1992) exhibits various intuitive and theoretical advantages over many other methods for unsupervised redundancy reduction. So far, however, there have not been any serious practical applications of PM. In this paper, we apply semilinear PM to static real world images and find that without a teacher and without any significant preprocessing, the system automatically learns to generate distributed representations based on well-known feature detectors, such as orientation-sensitive edge detectors and off-center–on-surround detectors, thus extracting simple features related to those considered useful for image preprocessing and compression.