Method and apparatus for accelerating data processing in neural network

1. An apparatus for accelerating data processing on a neural network, the apparatus comprising:

a control unit configured to quantize data by at least one method according to a characteristic of data calculated at a node forming at least one layer constituting the neural network, and to separately perform calculation at the node according to the quantized data; and

memory configured to store the quantized data,

wherein the data comprises an activation adapted to be output data at the node and a weight adapted to be an intensity at which the activation is reflected into the node; and

the control unit classifies the activation as any one of an outlier activation larger than a preset number of bits and a normal activation equal to or smaller than the preset number of bits based on a number of bits representing a quantized activation.

2. The apparatus of claim 1 , wherein the control unit determines the quantization method, to be applied to the data, based on a value of the data on the neural network.

3. The apparatus of claim 2 , wherein the control unit varies a number of bits of data attributable to the quantization based on an error between the data and the quantized data.

4. The apparatus of claim 1 , wherein the control unit separates the outlier activation and the normal activation from each other, and separately performs calculation on them.

5. The apparatus of claim 1 , wherein the control unit stores at least one of the normal activation and the weight in a buffer.

6. The apparatus of claim 1 , wherein the control unit, when an outlier weight, which is a weight larger than a preset number of bits, is present in the buffer, separates the outlier weight into a preset number of upper bits and remaining bits and then performs calculation.

7. A method for accelerating data processing on a neural network, the method comprising:

quantizing data by at least one method according to a characteristic of data calculated at a node forming at least one layer constituting the neural network;

storing the quantized data; and

separately performing calculation at the node according to the quantized data,

wherein the data comprises an activation adapted to be output data at the node and a weight adapted to be an intensity at which the activation is reflected into the node; and

the method comprises classifying the activation as any one of an outlier activation larger than a preset number of bits and a normal activation equal to or smaller than the preset number of bits based on a number of bits representing a quantized activation.

8. The method of claim 7 , wherein quantizing the data comprises determining the quantization method, to be applied to the data, based on a value of the data on the neural network.

9. The method of claim 8 , wherein determining the quantization method comprises varying a number of bits of data attributable to the quantization based on an error between the data and the quantized data.

10. The method of claim 7 , wherein separately performing the calculation comprises separating the outlier activation and the normal activation from each other and then performing calculation on them.

11. The method of claim 7 , wherein performing the calculation at the node comprises storing at least one of the normal activation and the weight in a buffer.

12. The method of claim 7 , further comprising, when an outlier weight, which is a weight larger than a preset number of bits, is present in the buffer, separating the outlier weight into a preset number of upper bits and remaining bits and then performing calculation.

13. A non-transitory computer-readable storage medium having stored thereon a program which performs the method set forth in claim 7 .

14. A computer program stored in a non-transitory computer-readable storage medium such that a neural network acceleration apparatus performs the method set forth in claim 7 .