Processor for accelerating convolutional operation in convolutional neural network and operating method thereof

A device for performing a convolutional operation in a convolutional neural network, the device comprising:

• a memory in which at least one program is stored;
• and
• a processor configured to perform a convolution operation in a convolutional neural network by executing the at least one program,
• wherein the processor is further configured to:
• obtain input activation data quantized with a first bit-width, from an input image, based on a first distance vector having a number of components that is the same as the number of bits of the first bit-width, wherein quantizing the input activation data is performed based on a sum of values obtained by multiplying values of components of the first distance vector by a set of preset values;
• obtain weight data quantized with a second bit-width, representing a value of a parameter learned through the convolutional neural network, based on a second distance vector having a number of components that is the same as the number of bits of the second bit-width, wherein quantizing the weight data is performed based on a sum of values obtained by multiplying values of components of the second distance vector by a set of preset values;
• binarize each of the input activation data and the weight data to obtain a binarization input activation vector and a binarization weight vector, respectively, wherein the processor binarizes by transforming data into vectors of binary numbers;
• perform an inner operation based on binary operation values, derived from components of the binarization input activation vector and the binarization weight vector, and the first and second distance vectors, wherein the processor performs the inner operation by:
• multiplying each of a result of a predetermined operation using the binary operation values by a corresponding component of the first distance vector and a corresponding component of the second distance vector, and
• summing the multiplication result values;
• and
• store a result obtained by the inner operation as output activation data.

A computer program stored on a non-transitory machine readable medium for executing a method of performing a convolution operation in a convolutional neural network in combination with a processor, the method comprising:

• loading an address to store output activation data representing a result obtained by performing an inner operation with respect to input activation data quantized with a first bit-width based on a first distance vector having a number of components that is the same as the number of bits of the first bit-width, wherein quantizing the input activation data is performed based on a sum of values obtained by multiplying values of components of the first distance vector by a set of preset values, and weight data quantized with a second bit-width based on a second distance vector having a number of components that is the same as the number of bits of the second bit-width, wherein quantizing the weight data is performed based on a sum of values obtained by multiplying values of components of the second distance vector by a set of preset values;
• loading a binarization input activation vector in which components in the input activation data are binarized and a binarization weight vector in which components in the weight data are binarized, wherein binarizing comprises transforming data into vectors of binary numbers;
• performing an inner operation based on binary operation values, derived from components of the binarization input activation vector and the binarization weight vector, and the first and second distance vectors, wherein the inner operation is performed by:
• multiplying each of a result of a predetermined operation using the binary operation values by a corresponding component of the first distance vector and a corresponding component of the second distance vector;
• and
• summing the multiplication result values;
• and
• storing the result obtained by the inner operation as the output activation data at the loaded address.