Multiply-accumulate system and multiply-accumulate method
A multiply-accumulate system ( 1 ) includes: a statistic calculation unit ( 111 ) that executes a standardization calculation for an input signal; and a multiply-accumulate device ( 10 ) that executes multiplication-accumulation based on the standardized input signal.
1. A multiply-accumulate system comprising:
a memory storing program code; and
a processor configured to execute the program code to perform operations comprising:
receiving an input signal including a plurality of pixel values;
executing a standardization calculation for each of the pixel values in the input signal to produce a standardized input signal;
multiplying at least one of the pixel values in the standardized input signal by a coefficient to produce a revised standardized input signal having respective pixel values corresponding to a predetermined range;
clipping the revised standardized input signal to produce a clipped input signal, the clipping ensuring that each of the pixel values in the clipped input signal are within the predetermined range; and
executing multiplication-accumulation on the clipped input signal.
2. The multiply-accumulate system according to claim 1 , wherein the standardization calculation is based on an average and standard deviation of the pixel values of the input signal.
3. The multiply-accumulate system according to claim 1 , wherein the memory holds learning data used for learning corresponding to executing the multiplication-accumulation, and the standardization calculation is based on the learning data.
4. The multiply-accumulate system according to claim 1 , wherein the predetermined range corresponds to an input value range of executing the multiplication-accumulation.
5. A multiply-accumulate method comprising:
receiving an input signal including a plurality of pixel values;
executing a standardization calculation for each of the pixel values in the input signal to produce a standardized input signal;
multiplying at least one of the pixel values in the standardized input signal by a coefficient to produce a revised standardized input signal having respective pixel values corresponding to a predetermined range;
clipping the revised standardized input signal to produce a clipped input signal, the clipping ensuring that each of the pixel values in the clipped input signal are within the predetermined range; and
executing multiplication-accumulation on the clipped input signal.
6. The multiply-accumulate method according to claim 5 , wherein the standardization calculation is based on an average and standard deviation of the pixel values of the input signal.
7. The multiply-accumulate method according to claim 5 , wherein a memory holds learning data used for learning corresponding to executing the multiplication-accumulation, and the standardization calculation is based on the learning data.
8. The multiply-accumulate method according to claim 5 , wherein the predetermined range corresponds to an input value range of executing the multiplication-accumulation.
9. A non-transitory computer readable medium storing program code for multiply-accumulation, the program code being executable by a processor to perform operations comprising:
receiving an input signal including a plurality of pixel values;
executing a standardization calculation for each of the pixel values in the input signal to produce a standardized input signal;
multiplying at least one of the pixel values in the standardized input signal by a coefficient to produce a revised standardized input signal having respective pixel values corresponding to a predetermined range;
clipping the revised standardized input signal to produce a clipped input signal, the clipping ensuring that each of the pixel values in the clipped input signal are within the predetermined range; and
executing multiplication-accumulation on the clipped input signal.
10. The non-transitory computer readable medium according to claim 9 , wherein the standardization calculation is based on an average and standard deviation of the pixel values of the input signal.
11. The non-transitory computer readable medium according to claim 9 , wherein a memory holds learning data used for learning corresponding to executing the multiplication-accumulation, and the standardization calculation is based on the learning data.
12. The non-transitory computer readable medium according to claim 9 , wherein the predetermined range corresponds to an input value range of executing the multiplication-accumulation.