Methods for long-range federated edge learning with chirp-based over-the-air computation
The disclosure deals with a system and method for circularly-shifted chirp (CSC)-based majority vote (MV) (CSC-MV), a low-peak-to-mean envelope power ratio (PMEPR) over-the-air computation (OAC) scheme, to achieve long-range federated edge learning (FEEL). The presently disclosed OAC approach maps the votes (i.e., the sign of the local gradients) from the edge devices (EDs) to linear CSCs constructed with the discrete Fourier transform (DFT)-spread orthogonal frequency division multiplexing (OFDM) (DFT-s-OFDM). At the edge server (ES), the MV was calculated non-coherently with an energy detector. We compare our presently disclosed scheme with one-bit broadband digital aggregation (OBDA) and show that the output-power back-off (OBO) requirement of the transmitters with an adjacent-channel-leakage ratio (ACLR) constraint for the presently disclosed scheme is significantly lower than the one with OBDA. We demonstrate that CSC-MV outperforms OBDA in terms of test accuracy for both homogeneous and heterogeneous data distributions, even when the power control is not ideal.
1 . A low peak-to-mean envelope power ratio (PMEPR) over-the-air computation (OAC) methodology for long-range federated edge learning (FEEL) without using channel state information (CSI) at a plurality of edge devices (EDs) or at an edge server (ES), comprising: a distributed machine-learning model to be trained with the update vectors received at the edge server (ES) as transmitted from the plurality of edge devices (EDs); one or more processors; and one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising: transmitting local update vector as votes in a form of circularly-shifted chirps based on the sign of the local update vector's elements, from each respective of the plurality of edge devices (EDs) via a wireless multiple access channel, receiving the superposed said local updates vectors at the ES, detecting the majority vote (MV) for each element of the local update vector at the ES with a circularly-shifted chirp (CSC)-based energy detector, and inputting the MVs into the machine-learning model for the model parameters to be updated.
2 . The methodology according to claim 1 , wherein the votes from the EDs include the signs of the local gradients.
3 . The methodology according to claim 2 , wherein mapping includes mapping the votes to linear CSCs constructed with pulse-position modulation (PPM) symbols constructed with discrete Fourier transform (DFT)-spread orthogonal frequency division multiplexing (OFDM) (DFT-s-OFDM), whereby the DFT-s-OFDM carry the gradient information with chirps.
4 . The methodology according to claim 3 , wherein at the ES, aggregated MV is calculated non-coherently.
5 . The methodology according to claim 3 , wherein chirp signals are synthesized through DFT-s-OFDM through the choice of frequency-domain spectral shaping (FDSS) coefficients.
6 . The methodology according to claim 3 , wherein t k,m comprises a linear combination of linear CSCs where the amount of frequency sweep for each CSC is D/T chirp for symbol duration T chirp where t k,m is the mth transmitted baseband signal in discrete time for the kth ED.
7 . The methodology according to claim 3 , wherein receiving operations include the ES detecting MV with a non-coherent detector by taking delay spread and synchronization errors into account to eliminate need for truncated-channel inversion (TCI) at the EDs.
8 . The methodology according to claim 1 , wherein the machine learning model comprises artificial intelligence technology over wireless or sensor networks, 5G or higher, 6G wireless standardization, or IEEE 802.11 Wi-Fi.
9 . The methodology according to claim 3 , wherein MV for the ith gradient at the MV v(n) i for the kth ED at the nth communication round between at least one ED and the ES comprises:
v
i
(
n
)
=
Δ
sign
(
∑
k
=
1
K
sign
(
g
~
k
,
i
(
n
)
)
)
=
sign
(
∑
k
=
1
K
g
_
k
,
i
(
n
)
)
.
10 . The methodology according to claim 1 , wherein the machine-learning model is training to learn the task of handwritten digit recognition.
11 . The methodology according to claim 1 , wherein the machine-learning model comprises a convolution neural network with multiple convolutional layers, with each convolutional layer followed by a batch normalization layer and rectified-linear unit (ReLU) activation following each of them.
12 . The methodology according to claim 11 , wherein the multiple convolutional layers each have a plurality of filters, and a fully connected layer with plural units and a softmax layer are used after one of the ReLU.
13 . The methodology according to claim 1 , further comprising detector operations at the ES that the detector compares the energies on two adjacent CSCs to determine a gradient vector, with a guard period in the time domain between two adjacent CSCs.
14 . A low peak-to-mean envelope power ratio (PMEPR) over-the-air computation (OAC) system for long-range federated edge learning (FEEL) without using channel state information (CSI) at a plurality of edge devices (EDs) or at an edge server (ES), comprising: a machine-learning model training to process update vectors received at the edge server (ES) as transmitted from the plurality of edge devices (EDs); one or more processors; and one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising: transmitting local update vectors as votes in a form of circularly-shifted chirps based on a sign of the local update vector's elements over selected multiple subcarriers from each respective of the plurality of edge devices (EDs) via a wireless multiple access channel, receiving the superposed said local updates vectors at the ES, mapping the majority vote (MV) for each element of the local update vector at the ES with a circularly-shifted chirp (CSC)-based energy detector, and inputting the MVs into the machine-learning model for the model parameters to be updated, transmitting local update vector as votes in the form of circularly-shifted chirps based on the sign of the vector's elements, from each respective of the plurality of edge devices (EDs) via a wireless multiple access channel, receiving the superposed local updates at the ES, detecting the majority vote (MV) for each element of the update vector at the ES with a circularly-shifted chirp (CSC)-based energy detector, and inputting the MVs into the machine-learning model for the model parameters to be updated.
15 . The system according to claim 14 , further including transmitters for transmitting the votes over multiple orthogonal subcarriers; and wherein, to determine the majority vote (MV) for each element of the update vector, the one or more processors are further programmed to perform mapping the votes to linear CSCs constructed with pulse-position modulation (PPM) symbols constructed with discrete Fourier transform (DFT)-spread orthogonal frequency division multiplexing (OFDM) (DFT-s-OFDM), whereby the DFT-s-OFDM carry the gradient information with chirps.
16 . The system according to claim 14 , wherein, to perform receiving operations, the one or more processors are further programmed to include the ES detecting MV with a non-coherent detector by taking delay spread and synchronization errors into account to eliminate need for truncated-channel inversion (TCI) at the EDs.
17 . The system according to claim 14 , wherein the machine-learning model comprises a convolution neural network with multiple convolutional layers, with each convolutional layer followed by a batch normalization layer and rectified-linear unit (ReLU) activation following each of them.
18 . The system according to claim 14 , wherein the system is configured to operate over wireless or sensor networks, 5G or higher, 6G wireless standardization, or IEEE 802.11 Wi-Fi.