GPU-based artificial intelligence system using channel-level architecture search for deep neural network

An electric apparatus for an artificial intelligence system comprising:

• an input interface configured to receive a first channel group including a plurality of channels;
• a memory configured to store at least one instruction;
• and
• a processor including a plurality of graphic processing units (GPUs) and configured to control each of the plurality of GPUs corresponding to each of a plurality of cells including a plurality of nodes,
• wherein
• a first graphic processing unit (GPU), among the plurality of GPUs, is configured to receive, via the input interface, a plurality of first channels in units of channel among a plurality of channels included in the first channel group,
• the first GPU is configured to obtain the respective first parameter weight indexes corresponding to weights of the respective first parameters by calculating the respective first parameters corresponding to the received plurality of first channels with the received first channels,
• the first GPU is configured to generate and output a second channel group by combining one first channel and the other channel of the plurality of first channels,
• a second GPU, among the plurality of GPUs, is configured to receive, via the input interface, a plurality of second channels in units of channel among a plurality of channels included in the generated second channel group,
• the second GPU is configured to obtain the respective second parameter weight indexes corresponding to weights of the respective second parameters by calculating the respective second parameters corresponding to the received plurality of second channels with the received second channels,
• the first GPU is configured to identify a first parameter having the highest first parameter weight index among the plurality of first parameters as a first architecture parameter, and identify a second parameter having the highest second parameter weight index among the plurality of second parameters as a second architecture parameter,
• the first GPU is configured to remove the first parameters other than the identified first architecture parameter among the plurality of first parameters, and remove the second parameters other than the identified second architecture parameter among the plurality of second parameters,
• the first GPU is configured to receive only the first channel corresponding to the identified first architecture parameter among the plurality of channels included in the first channel group,
• the first GPU is configured to modify an error of weight of the first architecture parameter by calculating the first channel corresponding to the first architecture parameter with the first architecture parameter,
• the first GPU is configured to generate and output a second channel group by combining the first channel corresponding to the first architecture parameter with the other channel,
• the second GPU is configured to receive only the second channel corresponding to the identified second architecture parameter among the plurality of channels included in the second channel group, and
• the second GPU is configured to modify an error of weight of the second architecture parameter by calculating the second channel corresponding to the second architecture parameter with the second architecture parameter,
• wherein
• a first node of the plurality of nodes includes a first plurality of operators, and
• a first operator of the first plurality of operators is configured to receive the number of channels corresponding to the number of first parameters included in the first operator from each of the channels of the first channel group.