Both from the clinical and mathematical perspectives, symptom recognition has received less attention than disease recognition. To redress this balance, it is imperative that multidimensional models are constructed for each and all mental symptoms. This paper offers one such model for ‘hallucinations,’ and a set of prototypical data comparing the performance of pattern recognition techniques (cluster and discriminant analyses) and neural networks (Kohonen and backpropagation). It is concluded that multidimensional models are less wasteful of information than (current) categorial ones. Because of this and of the fact that symptom structure is likely to be ‘isomorphic’ with the brain region where the corresponding signal is generated, it is recommended that multidimensional models are preferentially used in neurobiological research.