[2P3-J-2] Machine learning: advances in neural networks

[2P3-J-2-01] Proposal of optimization method of local receptive field using gradient descent

〇So Tazume¹, Masanao Ochi¹, Ichiro Sakata¹, Junichiro Mori¹ (1. The University of Tokyo)

Keywords:Neural Architecture Search(NAS), receptive field , Generalization AI

In this research, we try to design a method to search the optimum model according to individual data.In the existing structure search algorithm, the applicable types of data are limited.In this study, we will input the receptive field of the convolutional layer , It is possible to apply various types of data to the convolutional neural network.In concrete terms, by interpreting the receptive field as an index that associates inputs and weights in the convolutional layer We designed a new layer expressing the receptive field in a matrix and also made it possible to learn receptive field by gradient by relaxing its matrix to continuous value.The result of the experiment The proposed method for the data whose data structure is unknown By using the method proposed in this paper, it is expected that the spread of IoT and the sensor Relative big data that is expected to increase in the improvement of technology, the neural network is expected to be effectively applied.