Tohru Nitta

In this paper, we attempt to implement neural networks (NN) that do not have the critical points based on the hierarchical structure. First, we derive the sufficient conditions for the real-valued and complex-valued NNs not to have the critical points based on the hierarchical structure. Next, we construct the real-valued and complex-valued NNs that do not have the critical points based on the hierarchical structure by decomposing the high-dimensional NNs into the lower-dimensional NNs equivalent to the original high-dimensional NNs, as an application of the sufficient conditions.

We propose an emotion-memory model on neural networks which combines an emotional function of amygdala and relational-association of hippocampus. Amygdala generates emotion such as like/dislike from sensory inputs according to its inherent value judgment function, while hippocampus learns/recognizes an object from the same sensory inputs. Simulation result shows that this model can generate emotion for the object on hippocampus, by connecting amygdala and hippocampus through emotion.

We propose an associative memory model on hippocampal three-dimensional neural network model which is based on anatomical and physiological studies. Our model consists of following three assumptions, which are DG encodes input stimuli to orthogonal patterns, CAl forms relational association from the orthogonal pattern thorough its inter-lamellar connections, and subiculum decodes relations to restore the input stimuli patterns. As a result, our model realized essential function of associative memories, such as one-to-many, many-to-many and many-to-one associations, while concept forming and temporal association are also realized.

In this paper we propose a memory system which employs emotion as a control signal for memory. This Emotional Memory system can learn concepts without losing memory instances, and frequently accessed memories become easily associated.