Secure, accurate and fast neural network inference by replacing at least one non-linear activation channel
A method of a server device is provided. The method of a server device includes retrieving a prediction input and a prediction setting, replacing at least one non-linear activation channel in a neural network with at least one replacement channel based on the received prediction setting, generating a prediction based on the received prediction input based on the neural network with the at least one replacement channel, and outputting the generated prediction.
1 . A method of a server device, the method comprising:
receiving a prediction input and a prediction setting comprising an accuracy threshold;
providing a neural network operating on the server device and comprising a plurality of channels, the neural network being configured to generate a prediction based on the received prediction input;
based on the received prediction setting, selecting at least one non-linear activation channel among the plurality of channels that uses a non-linear activation and has a contribution to prediction accuracy of the neural network below the accuracy threshold;
replacing the at least one non-linear activation channel with at least one replacement channel having at least one approximated linear polynomial that utilizes fewer resources of the server device than the non-linear activation;
generating a prediction based on the received prediction input using the neural network with the at least one replacement channel; and
outputting the generated prediction.
2 . The method of claim 1 , wherein the prediction setting comprises a prediction generation speed.
3 . The method of claim 1 , wherein the prediction setting comprises a prediction generation accuracy.
4 . The method of claim 1 , wherein the selecting the at least one non-linear activation channel comprises determining a replacement ratio based on the prediction setting.
5 . The method of claim 4 , wherein the selecting the at least one non-linear activation channel further comprises selecting a percentage of non-linear activation channels in a layer of the neural network based on the determined replacement ratio.
6 . The method of claim 5 , wherein the replacing the at least one non-linear activation channel further comprises determining a replacement option based on the prediction setting.
7 . The method of claim 6 , wherein the replacing the at least one non-linear activation channel further comprises replacing the percentage of non-linear activation channels in the layer of the neural network with a polynomial approximation function with a degree that is determined based on the replacement option.
8 . The method of claim 1 , wherein the prediction setting is generated based on a user input to an interface of a client device.
9 . The method of claim 1 , wherein the at least one non-linear activation channel comprises any one or any combination of a rectified linear unit (ReLU) activation layer, a sigmoid function activation layer, a tangent function activation layer, a softmax function activation layer, and an exponential linear unit (ELU) layer.
10 . A method of a client device, the method comprising:
obtaining a prediction input;
generating a prediction setting comprising an accuracy threshold, based on a user input to an interface of the client device;
inputting the prediction input to a neural network comprising a plurality of channels, the neural network being configured to generate a prediction based on the prediction input;
inputting the prediction setting to modify the neural network; and
receiving a prediction generated from the modified neural network,
wherein the neural network is modified by:
based on the prediction setting, selecting at least one non-linear activation channel among the plurality of channels that uses a non-linear activation and has contribution to prediction accuracy of the neural network below the accuracy threshold; and
replacing the at least one non-linear activation channel with at least one replacement channel, and
wherein the at least one replacement channel has at least one approximated linear polynomial that utilizes fewer resources than the non-linear activation.
11 . The method of claim 10 , wherein the prediction setting comprises a prediction generation speed.
12 . The method of claim 10 , wherein the prediction setting comprises a prediction generation accuracy.
13 . The method of claim 10 , wherein the at least one non-linear activation channel is selected by determining a replacement ratio based on the prediction setting and selecting a percentage of non-linear activation channels in a layer of the neural network based on the determined replacement ratio.
14 . The method of claim 13 , wherein the at least one non-linear activation channel is replaced by determining a replacement option based on the prediction setting and replacing the percentage of non-linear activation channels in the layer of the neural network with a polynomial approximation function with a degree that is determined based on the replacement option.
15 . The method of claim 10 , wherein the at least one non-linear activation channel comprises any one or any combination of a rectified linear unit (ReLU) activation layer, a sigmoid function activation layer, a tangent function activation layer, a softmax function activation layer, and an exponential linear unit (ELU) layer.
16 . A server device, comprising:
a planner;
a neural network comprising a plurality of channels, the neural network being configured to generate a prediction;
at least one processor; and
memory that stores instructions that, when executed, cause the at least one processor to:
receive a prediction input and a prediction setting comprising an accuracy threshold;
select, with the planner, at least one non-linear activation channel among the plurality of channels that uses a non-linear activation and has a contribution to prediction accuracy of the neural network below the accuracy threshold, based on the received prediction setting;
replace, with the planner, the at least one non-linear activation channel in with at least one replacement channel having at least one approximated linear polynomial that utilizes fewer resources of the server device than the non-linear activation;
generate a prediction based on the received prediction input using the neural network with the at least one replacement channel; and
output the generated prediction.
17 . The server device of claim 16 , wherein the prediction setting comprises a prediction generation speed.
18 . The server device of claim 16 , wherein the prediction setting comprises a prediction generation accuracy.
19 . The server device of claim 16 , wherein the at least one non-linear activation channel is selected by determining a replacement ratio based on the prediction setting and selecting a percentage of non-linear activation channels in a layer of the neural network based on the determined replacement ratio.
20 . The server device of claim 19 , wherein the at least one non-linear activation channel is replaced by determining a replacement option based on the prediction setting and replacing the percentage of non-linear activation channels in the layer of the neural network with a polynomial approximation function with a degree that is determined based on the replacement option.