Learning with recurrent neural networks

This thesis examines so-called folding neural networks as a mechanism for machine learning. Folding networks form a generalization of partial recurrent neural networks such that they are able to deal with tree structured inputs instead of simple linear lists. In particular, they can handle classical formulas { they were proposed originally for this purpose. After a short explanation of the neural architecture we show that folding networks are well suited as a learning mechanism in principle. This includes three parts: the proof of their universal approximation ability, the aspect of information theoretical learnability, and the examination of the complexity of training. Approximation ability: It is shown that any measurable function can be approximated in probability. Explicit bounds on the number of neurons result if only a nite number of points is dealt with. These bounds are new results in the case of simple recurrent networks, too. Several restrictions occur if a function is to be approximated in the maximum norm. Afterwards, we consider brie y the topic of computability. It is shown that a sigmoidal recurrent neural network can compute any mapping in exponential time. However, if the computation is subject to noise almost the capability of tree automata arises. Information theoretical learnability: This part contains several contributions to distribution dependent learnability: The notation of PAC and PUAC learnability, consistent PAC/ PUAC learnability, and scale sensitive versions are considered. We nd equivalent characterizations of these terms and examine their respective relation answering in particular an open question posed by Vidyasagar. It is shown at which level learnability only because of an encoding trick is possible. Two approaches from the literature which can guarantee distribution dependent learnability if the VC dimension of the concept class is in nite are generalized to function classes: The function class is strati ed according to the input space or according to a so-called luckiness function which depends on the output of the learning algorithm and the concrete training data. Afterwards, the VC, pseudo-, and fat shattering dimension of folding networks are estimated: We improve some lower bounds for recurrent networks and derive new lower bounds for the pseudodimension and lower and upper bounds for folding networks in general. As a consequence, folding architectures are not distribution independent learnable. Distribution dependent learnability can be guaranteed. Explicit bounds on the number of examples which guarantee valid generalization can be derived using the two approaches mentioned above. We examine in which cases these bounds are polynomial. Furthermore, we construct an explicit example for a learning scenario where an exponential number of examples is necessary. Complexity: It is shown that training a xed folding architecture with perceptron activation function is polynomial. Afterwards, a decision problem, the so-called loading problem, which is correlated to neural network training is examined. For standard multilayer feed-forward networks the following situations turn out to be NP-hard: Concerning the perceptron activation function, a classical result from the literature, the NP-hardness for varying input dimension, is generalized to arbitrary multilayer architectures. Additionally, NP-hardness can be found if the input dimension is xed but the number of neurons may vary in at least two hidden layers. Furthermore, the NP-hardness is examined if the number of patterns and number of hidden neurons are correlated. We nish with a generalization of the classical NP result as mentioned above to the sigmoidal activation function which is used in practical applications.