Device for training neural networks in terms of hardware and energy requirements

摘要

Device for training a neural network (1) with an arrangement of neurons or other processing units, the architecture of which (1a) by architectural parametersa is characterized and where the behaviour of the neural network (1) is characterized by neuron parameters w which are associated with the neurons or other processing units present in accordance with the architecture, the device includes a machine-readable storage medium on which commands are stored; which, when running through a computer, have caused the computer to perform a procedure with the following steps:(a) the architectural parameters are initialized (110);Training data (2a) are provided (120);(a) the neural network (1) is established on the basis of the architectural parameters i, a (130);Training data (2a) are represented by the neural network (1) on outputs (3) on outputs (140);An expenditure quality measure L* is evaluated (150) which evaluates expenditure (3) in relation to the task assigned to the neural network (1);(i) the architectural parameters (i), (a) and (w) the neuron parameters (w) are optimized at least to the objectives (180);o further processing of training data (2a) through the neural network (1) improves the output quality measure L* and additionallyo a hardware effort for the implementation of architecture characterised by the architecture parameters i(a) on a hardware platform and/or energy consumption in the operation of the neural network on that hardware platform is reduced and/or meets at least one specified ancillary condition.