Quantifying inductive bias: AI learning algorithms and Valiant's learning framework
Artificial Intelligence
Hierarchical mixtures of experts and the EM algorithm
Neural Computation
Hi-index | 0.00 |
The goal of this research is to integrate aspects of artificial neural networks (ANNs) with symbolic machine learning methods in a probabilistic reasoning framework. Improved understanding of the semantics of neural nets supports principled integration efforts between seminumerical (so-called "subsymbolic") and symbolic intelligent systems. My dissertation focuses on learning of spatiotemporal (ST) sequences. In recent work, I have investigated architectures for modeling of ST sequences, and dualities between Bayesian networks and ANNs that expose their probabilistic and information theoretic foundations. In addition, I am developing algorithms for automated construction of Bayesian networks (and hybrid models); metrics for comparison of Bayesian networks across architectures; and a quantitative theory of feature construction (in the spirit of the PAC formalism from computational learning theory) for this learning environment. (Haussler 1988) Such methods for pattern prediction will be useful for building advanced knowledge based systems, with diagnostic applications such as intelligent monitoring tools.