Wireless sensing using classifier probing and refinement
Wireless sensing using classifier probing and refinement is described. In one example, a described method comprises: computing output analytics by a classifier based on input data constructed based on raw measurement data; mapping each output analytics to a respective mapped outcome; identifying at least one reference input data each associated with a reference output analytics and a reference mapped outcome; each reference input data being one of the input data for the classifier for which a respective reference outcome is available and is different from the reference mapped outcome; and for each reference input data for the classifier: computing a respective plurality of perturbed output analytics by the classifier, generating selected perturbed input data based on the plurality of perturbed output analytics; and re-training the classifier based on the selected perturbed input data and the associated reference outcome.
1 . A method performed by a system for wireless sensing with classifier probing and refinement, comprising:
in a probing phase of the system:
obtaining a plurality of raw measurement data by a sensing device of the system,
processing the plurality of raw measurement data by a processor of the system to construct a plurality of input data for a classifier,
performing a classification by the processor using the classifier by inputting each of the plurality of input data to the classifier,
computing a plurality of output analytics by the classifier based on the plurality of input data, each output analytics computed by the classifier based on a respective input data,
mapping the plurality of output analytics to a plurality of mapped outcome, each output analytics mapped to a respective mapped outcome,
identifying at least one reference input data each associated with a reference output analytics and a reference mapped outcome, each reference input data being one of the plurality of input data for the classifier for which a respective reference outcome is available and is different from the reference mapped outcome, and
for each reference input data for the classifier:
constructing a respective plurality of perturbed input data for the classifier by perturbing the reference input data, each perturbed input data constructed based on a respective perturbation of the reference input data,
performing the classification using the classifier by inputting each of the plurality of perturbed input data to the classifier,
computing a respective plurality of perturbed output analytics by the classifier based on the plurality of perturbed input data, each perturbed output analytics computed by the classifier based on a respective perturbed input data,
mapping the plurality of perturbed output analytics to a plurality of perturbed mapped outcome, each perturbed output analytics mapped to a respective perturbed mapped outcome,
comparing each perturbed mapped outcome with the reference outcome associated with the reference input data which is different from the reference mapped outcome, and
when at least one perturbed mapped outcome deviates from the reference mapped outcome such that it is the same as the reference outcome, selecting at least one selected perturbed input data each being one of the respective plurality of perturbed input data associated with one of the at least one perturbed mapped outcome as selected perturbed input data for re-training the classifier; and
in a re-training phase of the system:
re-training the classifier based on each of the at least one selected perturbed input data and the associated reference outcome,
wherein for each input data being the reference input data or any perturbed input data:
the input data is a k-dimensional (k-D) matrix,
each component of the input data is a matrix element of the k-D matrix,
a first matrix element of the k-D matrix is a neighboring component near a second matrix element of the k-D matrix if a distance between coordinates of the first matrix element and the second matrix element is less than a threshold.
2 . The method of claim 1 , further comprising:
for a particular reference input data for the classifier in the probing phase:
constructing a number of perturbed input data by perturbing the particular reference input data with an increasing degree of perturbation; and
selecting a particular selected perturbed input data from the number of perturbed input data, wherein the particular selected perturbed input data has a minimum degree of perturbation to the particular reference input data such that the associated perturbed mapped outcome is the same as a particular reference outcome associated with the particular reference input data.
3 . The method of claim 2 , further comprising:
for the particular reference input data for the classifier in the probing phase:
performing a search among the number of perturbed input data in search of the particular selected perturbed input data, wherein the search comprises one of: full exhaustive search, sequential search with early termination, binary search, or another fast search.
4 . The method of claim 1 , further comprising:
for a reference input data for the classifier in the probing phase:
selecting, based on a particular selected perturbed input data, at least one supplementary selected perturbed input data that is in a neighborhood of the particular selected perturbed input data; and
in the re-training phase:
re-training the classifier further based on the at least one supplementary selected perturbed input data.
5 . The method of claim 4 , wherein:
the at least one supplementary selected perturbed input data comprise at least one of a number of perturbed input data that (1) are next to the particular selected perturbed input data and (2) have a minimal decrease or increase of degree of perturbation compared with the particular selected perturbed input data.
6 . The method of claim 1 , further comprising:
receiving, by the sensing device, a wireless signal from a wireless device through a wireless channel of a venue,
wherein the plurality of raw measurement data comprise a time series of channel information (TSCI) of the wireless channel obtained by the sensing device based on the received wireless signal.
7 . The method of claim 1 , further comprising:
mapping each output analytics to a respective mapped outcome based on at least one of: a nonlinear function of the output analytics, a vector quantization of the output analytics, or a thresholding of the output analytics.
8 . The method of claim 1 , further comprising:
for each of a plurality of perturbation types, constructing at least one perturbed input data by perturbing a reference input data based on the perturbation type.
9 . The method of claim 1 , further comprising:
constructing each input data for the classifier as a k-dimensional (k-D) matrix; and
constructing each perturbed input data as a k-dimensional matrix with a same size as the input data.
10 . The method of claim 9 , further comprising:
for each reference input data comprising a set of components, constructing each perturbed input data for the classifier based on the reference input data by:
marking a respective subset of the components of the reference input data to generate a respective subset of marked components, and
modifying the respective subset of marked components of the reference input data to construct the respective perturbed input data.
11 . The method of claim 10 , further comprising:
modifying a respective marked component of the reference input data by at least one of:
replacing the respective marked component by zero,
scaling a magnitude of the respective marked component, or
adding an additive distortion to the respective marked component.
12 . The method of claim 10 , further comprising:
modifying a respective marked component of the reference input data by at least one of:
adding an additive noise to the respective marked component; or
adding an additive distortion scaled by a scaling factor to the respective marked component, wherein the additive distortion is associated with a perturbation type and the scaling factor is associated with a perturbation size.
13 . The method of claim 10 , wherein marking the respective subset of the components of the reference input data to generate the respective subset of marked components comprises:
randomly selecting a component of the reference input data; and
selecting a respective number of neighboring components in the reference input data near the randomly selected component to be included in the respective subset of marked components.
14 . The method of claim 13 , wherein:
the respective number of selected neighboring components in the reference input data comprise a local neighborhood of matrix elements of the k-D matrix near the randomly selected component; and
the respective number of selected neighboring components in the reference input data are the respective subset of marked components.
15 . The method of claim 13 , wherein:
the respective number of selected neighboring components in the reference input data comprise a sub-matrix of matrix elements of the k-D matrix near the randomly selected component; and
the sub-matrix is a (k2)-dimensional (k2-D) sub-matrix comprising the randomly selected component, wherein k2<=k.
16 . The method of claim 1 , wherein the classifier comprises at least one of:
support vector machine (SVM), kernel-based SVM, neural network (NN), deep neural network (DN), convolutional NN (CNN), recurrent NN (RNN), long short-term memory network (LSTM), deep belief network (DBN), generative adversarial network (GAN), conditional GAN (cGAN), transformer, encoder-decoder, denoising autoencoder (DAE), neural radiance field, radial basis function network, multilayer perceptron (MLP), or self organizing map (SOM).
17 . A system for wireless sensing with classifier probing and refinement, comprising:
a sensing device configured to obtain a plurality of raw measurement data in a probing phase of the system; and
a classifier comprising a memory and a processor configured in the probing phase of the system to:
process the plurality of raw measurement data to construct a plurality of input data,
perform a classification using each of the plurality of input data as input to the classifier,
compute a plurality of output analytics based on the plurality of input data, each output analytics computed based on a respective input data,
map the plurality of output analytics to a plurality of mapped outcome, each output analytics mapped to a respective mapped outcome,
identify at least one reference input data each associated with a reference output analytics and a reference mapped outcome, each reference input data being one of the plurality of input data for the classifier for which a respective reference outcome is available and is different from the reference mapped outcome, and
for each reference input data for the classifier:
construct a respective plurality of perturbed input data for the classifier by perturbing the reference input data, each perturbed input data constructed based on a respective perturbation of the reference input data,
perform the classification using each of the plurality of perturbed input data as input to the classifier,
compute a respective plurality of perturbed output analytics based on the plurality of perturbed input data, each perturbed output analytics computed based on a respective perturbed input data,
map the plurality of perturbed output analytics to a plurality of perturbed mapped outcome, each perturbed output analytics mapped to a respective perturbed mapped outcome,
compare each perturbed mapped outcome with the reference outcome associated with the reference input data which is different from the reference mapped outcome, and
when at least one perturbed mapped outcome deviates from the reference mapped outcome such that it is the same as the reference outcome, select at least one selected perturbed input data each being one of the respective plurality of perturbed input data associated with one of the at least one perturbed mapped outcome as selected perturbed input data for re-training the classifier,
wherein the processor is configured to re-train the classifier based on each of the at least one selected perturbed input data and the associated reference outcome in a re-training phase of the system,
wherein for each input data being the reference input data or any perturbed input data:
the input data is a k-dimensional (k-D) matrix,
each component of the input data is a matrix element of the k-D matrix,
a first matrix element of the k-D matrix is a neighboring component near a second matrix element of the k-D matrix if a distance between coordinates of the first matrix element and the second matrix element is less than a threshold.
18 . The system of claim 17 , wherein the processor is further configured to:
for a particular reference input data for the classifier in the probing phase:
construct a number of perturbed input data by perturbing the particular reference input data with an increasing size of perturbation; and
select a particular selected perturbed input data from the number of perturbed input data, wherein the particular selected perturbed input data has a minimum size of perturbation to the particular reference input data such that the associated perturbed mapped outcome is the same as a particular reference outcome associated with the particular reference input data.
19 . The system of claim 18 , wherein the processor is further configured to:
for the particular reference input data for the classifier in the probing phase:
perform a search among the number of perturbed input data in search of the particular selected perturbed input data, wherein the search comprises one of: full exhaustive search, sequential search with early termination, binary search, or another fast search.
20 . The system of claim 17 , wherein the processor is further configured to:
for a reference input data for the classifier in the probing phase:
select, based on a particular selected perturbed input data, at least one supplementary selected perturbed input data that is in a neighborhood of the particular selected perturbed input data; and
in the re-training phase:
re-train the classifier further based on the at least one supplementary selected perturbed input data.
21 . The system of claim 17 , wherein the processor is further configured to:
map each output analytics to a respective mapped outcome based on at least one of: a nonlinear function of the output analytics, a vector quantization of the output analytics, or a thresholding of the output analytics.
22 . The system of claim 17 , wherein the processor is further configured to:
construct each input data for the classifier as a k-dimensional (k-D) matrix; and
construct each perturbed input data as a k-dimensional matrix with a same size as the input data.
23 . The system of claim 22 , wherein the processor is further configured to:
for each reference input data comprising a set of components, construct each perturbed input data for the classifier based on the reference input data by:
marking a respective subset of the components of the reference input data to generate a respective subset of marked components, and
modifying the respective subset of marked components of the reference input data to construct the respective perturbed input data.
24 . The system of claim 23 , wherein the processor is further configured to:
modify a respective marked component of the reference input data by at least one of:
replacing the respective marked component by zero,
scaling a magnitude of the respective marked component, or
adding an additive distortion to the respective marked component.
25 . The system of claim 23 , wherein the processor is further configured to:
modify a respective marked component of the reference input data by at least one of:
adding an additive noise to the respective marked component; or
adding an additive distortion scaled by a scaling factor to the respective marked component, wherein the additive distortion is associated with a perturbation type and the scaling factor is associated with a perturbation size.
26 . The system of claim 23 , wherein marking the respective subset of the components of the reference input data to generate the respective subset of marked components comprises:
randomly selecting a component of the reference input data; and
selecting a respective number of neighboring components in the reference input data near the randomly selected component to be included in the respective subset of marked components.
27 . A device for wireless sensing with classifier probing and refinement, comprising:
a receiver configured to obtain a plurality of raw measurement data in a probing phase of the device; and
a classifier comprising a memory and a processor configured in the probing phase of the device to:
process the plurality of raw measurement data to construct a plurality of input data,
perform a classification using each of the plurality of input data as input to the classifier,
compute a plurality of output analytics based on the plurality of input data, each output analytics computed based on a respective input data,
map the plurality of output analytics to a plurality of mapped outcome, each output analytics mapped to a respective mapped outcome,
identify at least one reference input data each associated with a reference output analytics and a reference mapped outcome, each reference input data being one of the plurality of input data for the classifier for which a respective reference outcome is available and is different from the reference mapped outcome, and
for each reference input data for the classifier:
construct a respective plurality of perturbed input data for the classifier by perturbing the reference input data, each perturbed input data constructed based on a respective perturbation of the reference input data,
perform the classification using each of the plurality of perturbed input data as input to the classifier,
compute a respective plurality of perturbed output analytics based on the plurality of perturbed input data, each perturbed output analytics computed based on a respective perturbed input data,
map the plurality of perturbed output analytics to a plurality of perturbed mapped outcome, each perturbed output analytics mapped to a respective perturbed mapped outcome,
compare each perturbed mapped outcome with the reference outcome associated with the reference input data which is different from the reference mapped outcome, and
when at least one perturbed mapped outcome deviates from the reference mapped outcome such that it is the same as the reference outcome, select at least one selected perturbed input data each being one of the respective plurality of perturbed input data associated with one of the at least one perturbed mapped outcome as selected perturbed input data for re-training the classifier,
wherein the processor is configured to re-train the classifier based on each of the at least one selected perturbed input data and the associated reference outcome in a re-training phase of the device,
wherein for each input data being the reference input data or any perturbed input data:
the input data is a k-dimensional (k-D) matrix,
each component of the input data is a matrix element of the k-D matrix,
a first matrix element of the k-D matrix is a neighboring component near a second matrix element of the k-D matrix if a distance between coordinates of the first matrix element and the second matrix element is less than a threshold.
28 . The device of claim 27 , wherein the processor is further configured to:
for a particular reference input data for the classifier in the probing phase:
construct a number of perturbed input data by perturbing the particular reference input data with an increasing degree of perturbation; and
select a particular selected perturbed input data from the number of perturbed input data, wherein the particular selected perturbed input data has a minimum degree of perturbation to the particular reference input data such that the associated perturbed mapped outcome is the same as a particular reference outcome associated with the particular reference input data.
29 . The device of claim 27 , wherein the processor is further configured to:
for a reference input data for the classifier in the probing phase:
select, based on a particular selected perturbed input data, at least one supplementary selected perturbed input data that is in a neighborhood of the particular selected perturbed input data; and
in the re-training phase:
re-train the classifier further based on the at least one supplementary selected perturbed input data.
30 . The device of claim 27 , wherein the processor is further configured to:
map each output analytics to a respective mapped outcome based on at least one of: a nonlinear function of the output analytics, a vector quantization of the output analytics, or a thresholding of the output analytics.