US 11,748,612 B2
Neural processing device and operation method thereof
Youngjoo Lee, Pohang-si (KR); Sunggu Lee, Pohang-si (KR); Minho Ha, Pohang-si (KR); and Younghoon Byun, Pohang-si (KR)
Assigned to POSTECH ACADEMY-INDUSTRY FOUNDATION, Gyeonsangbuk-do (KR)
Filed by Postech Academy-Industry Foundation, Pohang-si (KR)
Filed on Apr. 11, 2019, as Appl. No. 16/381,200.
Claims priority of application No. 10-2018-0116612 (KR), filed on Sep. 28, 2018.
Prior Publication US 2020/0104701 A1, Apr. 2, 2020
Int. Cl. G06N 3/08 (2023.01); G06N 5/04 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 5/04 (2013.01)] 16 Claims
OG exemplary drawing
 
1. A neural processing device comprising:
a processor configured to execute computer-readable instructions stored in a memory such that the processor is configured to cause the neural processing device to,
perform a preprocessing on input data to determine a noise characteristic of the input data,
select one of a plurality of neural networks and weights associated therewith based on the noise characteristic and with reference to a weight table, the weight table stored in a memory and including a noise characteristic field, a neural network field, and a weight address field, the plurality of neural networks being neural networks previously learned based on various input data including various noise characteristics, and
perform inference on the input data based on the selected weights corresponding to the selected neural network,
wherein the processor includes a first neural network used to perform the preprocessing, and the selected neural network used to perform the inference as a second neural network, and the first neural network includes fewer layers and fewer neurons than the second neural network,
wherein the noise characteristic comprises a noise type and a noise degree of the input data, and
wherein the processor is further configured to learn weights corresponding to each of the plurality of neural networks based on input data including noise corresponding to a combination of the noise type and the noise degree and perform the inference based on the learned weights.