A discrete data representation-supporting apparatus for executing the back-training of an artificial neural network. The apparatus comprises an instruction cache unit (1), a controller unit (2), a data access unit (3), an interconnected module (4), a master operation module (5), a plurality of slave operation modules (6), and a continuous-to-discrete conversion module (7). Using the apparatus can realize the back-training of a multi-layer artificial neural network. This apparatus emphasizes the support for discrete data, comprising the storage and operation of discrete data, and the conversion from continuous data to discrete data. Data of a weight value, a neuron value, etc. in a process of performing artificial neural network back calculation by means of using this apparatus all support a discrete-form representation or a continuous-form representation. The discrete data representation refers to a storage mode of representing data by means of a particular number. For example, four numbers, 00, 01, 10 and 11, can respectively stand for data, i.e. four numbers, -1, -1/8, 1/8 and 1. This kind of storage mode is different from a continuous storage mode, wherein binary 00/01/10/11 stand for four numbers, 0, 1, 2, and 3. The continuous-to-discrete conversion module provides a module for converting continuous data into discrete data.
展开▼
