US 12,282,840 B2
Method and apparatus with neural network layer contraction
Shih-Chii Liu, Zurich (CH); Bodo Rueckauer, Zurich (CH); and Tobi Delbruck, Zurich (CH)
Assigned to Samsung Electronics Co., Ltd., Suwon-si (KR); and University of Zurich, Zurich (CH)
Filed by Samsung Electronics Co., Ltd., Suwon-si (KR); and University of Zurich, Zurich (CH)
Filed on Jan. 10, 2020, as Appl. No. 16/739,543.
Claims priority of provisional application 62/791,237, filed on Jan. 11, 2019.
Claims priority of application No. 10-2019-0087099 (KR), filed on Jul. 18, 2019.
Prior Publication US 2020/0226451 A1, Jul. 16, 2020
Int. Cl. G06N 3/063 (2023.01); G06N 3/082 (2023.01); G06N 5/046 (2023.01)
CPC G06N 3/063 (2013.01) [G06N 3/082 (2013.01); G06N 5/046 (2013.01)] 24 Claims
OG exemplary drawing
 
1. A processor-implemented neural network method, the method comprising:
determining a reference sample among sequential input samples to be processed by a neural network, the neural network comprising an input layer, one or more hidden layers, and an output layer;
performing an inference process by calculating an output activation of the output layer based on operations of an activation function performed in the hidden layers corresponding to the reference sample input to the input layer;
determining layer contraction parameters that define an affine transformation relationship between the input layer and the output layer, for approximation of the inference process, the layer contraction parameters comprising a binary mask vector for performing activation masking by replacing an operation of the activation function performed in the hidden layers;
contracting, based on the layer contraction parameters, the neural network by performing activation masking using the binary mask vector to remove at least one of the hidden layers from the neural network to create a layer-contracted neural network; and
performing, using the layer-contracted neural network, in response to a value determined based on one or more pixel values of the reference sample and one or more pixel values of one or more input samples subsequent to the reference sample among the sequential input samples being greater than or equal to a threshold value, inference on the one or more input samples subsequent to the reference sample using affine transformation based on the layer contraction parameters determined with respect to the reference sample.