Method and electronic device for performing deep neural network operation

An electronic device for performing a convolutional neural network (CNN) operation, the electronic device comprising:

• an accelerator;
• a memory storing one or more instructions;
• and
• at least one processor configured to execute the one or more instructions stored in the memory to:
• perform a first modified operation at a first convolution layer of the CNN, wherein the first modified operation achieves an effect corresponding to a space-to-depth transformation operation, and comprises:
• modify a kernel of the first convolution layer to obtain a first modified kernel, and
• obtain a channel-extended output feature map directly from an input feature map of the first convolution layer by applying a convolution operation using the first modified kernel;
• perform a neural network operation corresponding to layers between a first layer and a second layer as a channel-extended neural network operation, wherein the second layer is a layer of a plurality of layers of the CNN after the first layer;
• perform a second modified operation at a second convolution layer of the CNN, wherein the second modified operation achieves an effect corresponding to a depth-to-space transformation operation, and comprises:
• modify a kernel of the second convolution layer to obtain a second modified kernel, and
• obtain a channel-reduced output feature map, directly from the channel-extended feature map input into the second convolution layer by applying a transposed convolution operation using the second modified kernel;
• and
• output a result of the CNN operation.

A non-transitory computer-readable recording medium having recorded thereon a program for executing, on a computer, a method by which an electronic device performs a convolutional neural network (CNN) operation, the method comprising:

• performing a first modified operation at a first convolution layer of the CNN, wherein the first modified operation achieves an effect corresponding to a space-to-depth transformation operation, and comprises:
• modifying a kernel of the first convolution layer to obtain a first modified kernel, and
• obtaining a channel-extended output feature map directly from an input feature map of the first convolution layer by applying a convolution operation using the first modified kernel;
• performing a neural network operation corresponding to layers between a first layer and a second layer as a channel-extended neural network operation, wherein the second layer is a layer of a plurality of layers of the CNN after the first layer;
• performing a second modified operation at a second convolution layer of the CNN, wherein the second modified operation achieves an effect corresponding to a depth-to-space transformation operation, and comprises:
• modifying a kernel of the second convolution layer to obtain a second modified kernel, and
• obtaining a channel-reduced output feature map, directly from a channel-extended feature map input into the second convolution layer by applying a transposed convolution operation using the second modified kernel;
• and
• outputting a result of the CNN operation.