CCS: Continuous Learning for Customized Incremental Wireless Sensing Services
Extracted source text · arXiv 2412.04821
% \documentclass[conference]{IEEEtran}
\documentclass[10pt,confrence]{IEEEtran}
\IEEEoverridecommandlockouts
% The preceding line is only needed to identify funding in the first footnote. If that is unneeded, please comment it out.
\usepackage{cite}
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{algorithmic}
\usepackage{graphicx}
\usepackage{textcomp}
\usepackage{xcolor}
\usepackage{subcaption} % 添加子图功能
\usepackage{booktabs}
\usepackage[normalem]{ulem}
\usepackage{bbding}
\usepackage{url}
\usepackage{svg}
\usepackage{booktabs}
\usepackage{multirow}
\usepackage{float}
\usepackage{hyperref}
\usepackage[numbers]{natbib}
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}
\usepackage{orcidlink}
\begin{document}
\title{CCS: Continuous Learning for Customized \\ Incremental Wireless Sensing Services
% {\footnotesize \textsuperscript{*}Note: Sub-titles are not captured in Xplore and
% should not be used
% }
% \thanks{Identify applicable funding agency here. If none, delete this.}
}
\author{Qunhang Fu\orcidlink{0009-0007-9789-0300},
Fei Wang\orcidlink{0000-0002-0750-6990},~\IEEEmembership{Member,~IEEE,}
Mengdie Zhu\orcidlink{0009-0007-3965-6700},\\
Han Ding\orcidlink{0000-0002-5274-7988},~\IEEEmembership{Member,~IEEE,}
Jinsong Han\orcidlink{0000-0001-5064-1955},~\IEEEmembership{Senior Member,~IEEE}
Tony Xiao Han,~\IEEEmembership{Senior Member,~IEEE}
\thanks{
Qunhang Fu(email: 3122358189@stu.xjtu.edu.cn), Fei Wang, and Mengdie Zhu are with the School of Software Engineering, Xi'an Jiaotong University, Xi'an China. Han Ding is with the School of Computer Science and Technology, Xi'an Jiaotong University, Xi'an China. Jinsong Han is with the College of Computer Science and Technology, Zhejiang University, Hangzhou, China. Tony Xiao Han is with the Wireless Technology Laboratory, Huawei Technologies Co. Ltd. Shenzhen, China. Fei Wang is the corresponding author (email: feynmanw@xjtu.edu.cn).
}
}
\markboth{IEEE Internet of Things Journal,~Vol.~00, No.~0, December~2024}%
{Shell \MakeLowercase{\textit{et al.}}: Bare Demo of IEEEtran.cls for IEEE Journals}
\maketitle
\begin{abstract}
Wireless sensing has made significant progress in tasks ranging from action recognition, vital sign estimation, pose estimation, etc. After over a decade of work, wireless sensing currently stands at the tipping point transitioning from proof-of-concept systems to the large-scale deployment. We envision a future service scenario where wireless sensing service providers distribute sensing models to users. During usage, users might request new sensing capabilities. For example, if someone is away from home on a business trip or vacation for an extended period, they may want a new sensing capability that can detect falls in elderly parents or grandparents and promptly alert them. In this paper, we propose CCS (continuous customized service), enabling model updates on users' local computing resources without data transmission to the service providers. To address the issue of catastrophic forgetting in model updates—where updating model parameters to implement new capabilities leads to the loss of existing capabilities—we design knowledge distillation and weight alignment modules. These modules enable the sensing model to acquire new capabilities while retaining the existing ones. We conducted extensive experiments on the large-scale XRF55 dataset across Wi-Fi, millimeter-wave radar, and RFID modalities to simulate scenarios where four users sequentially introduced new customized demands. The results affirm that CCS excels in continuous model services across all the above wireless modalities, significantly outperforming existing approaches like OneFi~\cite{onefi}.
% \thanks{Codes are available at this anonymous link for review:~\url{https://anonymous.4open.science/r/CCS-31EC/}}
\end{abstract}
\begin{IEEEkeywords}
continuous learning, wireless sensing service, Wi-Fi, millimeter-wave radar, RFID, human action recognition
\end{IEEEkeywords}
\section{Introduction}\label{sec:introduction}
After more than a decade of research, wireless sensing is now at a tipping point, transitioning from laboratory settings to large-scale deployment. Several developments have attested to this shift, including the consideration of integrated sensing and communication in the IEEE 802.11bf protocol standardization~\cite{802}, Google's development of Soli millimeter-wave radar chips for gesture recognition in the Pixel smartphones~\cite{soli}, and the emergence of numerous wireless sensing startups like ORIGIN Wireless. Perceiving the tipping point, we proactively delve into a prospective business schema for wireless sensing, in which two main participants are involved: wireless sensing service providers and users. Wireless sensing service providers initially distribute trained wireless sensing models to users, constituting their base service. Subsequently, during daily lives, users may encounter ongoing personalized requirements that the initial sensing models do not cater to. For instance, when a user needs to be away for an extended period, s/he may desire new sensing capability to detect falls in elderly parents or grandparents. If service providers can also respond to such customized needs, it will constitute another significant revenue stream.
\begin{figure}[t]
\centering
\includegraphics[width=1\linewidth]{figs/fig1.pdf}
\caption{We present CCS. With CCS, the service provider can not only provide base wireless sensing service for users but also provide incremental model service to meet users' continuous customized sensing demand under the premise that users' wireless sensing data is never transmitted to the service provider to ensure data privacy protection.}
\label{fig:ccs}
\end{figure}
The simplest solution would be for users to send their data to the service providers, whereupon providers retrain a model tailored to the new requirements and then return it to users. However, this approach risks compromising user data privacy. One privacy-preserving solution proposed is federated learning, a technique that allows users to transmit only parameters associated with the raw data to service providers, rather than the data itself~\cite{fl1,fl2,fl3}. However, some literature has pointed out that such parameters can be leveraged to reconstruct the original data~\cite{fl4,fl5}. Moreover, the above approaches may also raise network strain and computational burdens on the service providers, particularly as the user base expands significantly.
%This paper introduces CCS, a scheme enabling service providers to attain the capacity for delivering such value-added services. CCS operates under the reasonable assumption that user wireless sensing data is never transmitted to service providers due to data privacy concerns. The core idea behind CCS is straightforward, as depicted in Fig.~\ref{fig:ccs}, encompassing two distinct stages. In the first stage, the service provider distributes a wireless sensing model to users, enabling them to access wireless sensing services locally on their devices, even in situations where network connectivity is lost. The second stage addresses the scenario where the users encounter new sensing requirements. In this stage, CCS requires users to acquire wireless data that supports the new requirements and store it locally. The service provider then utilizes this data to update the sensing model with computing resources of users' devices, thereby enhancing the model with new capabilities to meet the users' new needs. We term these two stages as base model service and incremental model service, respectively. In addition, CCS enables the service provider to continuously provide the second phase of service, which we denominate as the continuously incremental model service, shown in Fig.~\ref{fig:ccs}.
This paper introduces CCS, a scheme enabling service providers to attain the capacity for delivering such continuous services. The design of CCS is predicated on a fundamental assumption: to safeguard data privacy, user wireless sensing data is never transmitted to the service provider. As depicted in Fig.~\ref{fig:ccs}, the implementation of this scheme is structured into three distinct stages: base model service, incremental model service, and continuously incremental model service. In the first stage, the service provider distributes a wireless sensing model to users, enabling them to access wireless sensing services locally on their devices, even in situations where network connectivity is lost.
As user requirements evolve, CCS facilitates the introduction of new sensing demands through the provision of new data by the users. At this stage, CCS requires users to acquire wireless data that supports the new requirements and store it locally. Service providers then leverage this data, harnessing the computational resources of the user's device, to update the model, thereby endowing it with the capability to meet the new demands. CCS supports not only singular model updates but also the ongoing evolution of model services, which we denominate as the continuously incremental model service, shown in Fig.~\ref{fig:ccs}.
The issue of catastrophic forgetting~\cite{cf1,cf2,cf3} often arises during model updates in incremental model services. This occurs when the model is updated with users' new data to meet new requirements, causing it to lose its existing capabilities. To maintain recognition accuracy for existing data categories while incorporating new ones, the service provider could select a subset of data from its repository, called exemplars\cite{ex1,ex2}, which would then be transmitted to the users' local devices for participation in model updating, as illustrated in Fig.\ref{fig:ccs}. Given that data is the vital asset for service providers, exemplars would be kept as compact as feasible. CCS leverages Herding techniques~\cite{Herding} to select exemplars, from the provider's data repository, and deliver the exemplars to users.
% Overall, the CCS framework not only enhances the flexibility and responsiveness of wireless sensing services but also provides effective technical support for protecting user data privacy. It demonstrates the potential for the evolution of wireless sensing services towards personalization and customization.
Furthermore, CCS initializes the current stage model $\mathcal{M}_{s}$ with model from the previous stage, $\mathcal{M}_{s} \gets \mathcal{M}_{s-1}$. $\mathcal{M}_{s}$ is updated with new user's data and exemplars from the provider. Meanwhile, the parameters of $\mathcal{M}_{s-1}$ are frozen as teacher network and distilled into $\mathcal{M}_{s}$ as knowledge of the old service, ensuring that $\mathcal{M}_{s}$ does not forget the previous service during updates. Moreover, CCS applies weight aligning techniques~\cite{wa} to balance the model's propensity towards the new and old services. As users continue to acquire new services, $\mathcal{M}_{s}$ will serve as the initialization for the model in the next stage, $\mathcal{M}_{s+1} \gets \mathcal{M}_{s}$, with these operations iteratively executed to achieve continuously incremental model service.
We use the large-scale XRF55 dataset~\cite{xrf55} to evaluate CCS. XRF55 encompasses 55 indoor action classes spanning human-computer interaction, human-human interaction, and human-object interaction, incorporating three wireless modalities: Wi-Fi, millimeter-wave radar, and RFID. During evaluation, we initially partition XRF55 into five subsets, each consisting of 15, 10, 10, 10, and 10 action categories, respectively. We posit that the service provider trains and distributes a base model using the subset containing 15 actions. Subsequently, the user progressively adds four incremental action recognition requirements of 10 classes each, employing the CCS methodology. We introduce new demand categories in various sequences across Wi-Fi, millimeter-wave radar, and RFID. Extensive experiments reveal that CCS effectively ensures model competence in recognizing newly added action categories while largely preserving its ability to identify previously learned action classes. To assess model worth, we adopt a novel metric – ACCN, the multiplication of the number of recognizable action categories and model accuracy, which can balance the model capacity and accuracy. Our validation demonstrates that this metric consistently improves with each incremental model service stage and significantly outperforms existing approaches, such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic}, OneFi~\cite{onefi}. We summarize the contributions of this paper as follows.
(1) We have contemplated a potential future business schema for wireless sensing services, encompassing service providers and users. For the first time, we have addressed the issue of how service providers might accommodate users' continuously emerging new requirements, proposing the CCS solution in this regard.
(2) The proposed CCS encompasses three key technical aspects: first, the selection of exemplars; second, model updating via knowledge distillation; and third, the alignment of weights for new and old services. Collectively, these techniques ensure that while CCS provides new services, it does not discard the existing services.
(3) We have conducted extensive comparative studies and ablation experiments on the WiFi, millimeter-wave radar, and RFID modalities within the XRF55 dataset. The results demonstrate that the proposed CCS effectively mitigates catastrophic forgetting issues encountered during continuous model service.
\section{RELATED WORK}\label{sec:related_work}
\subsection{Few-shot Learning}
Few-shot learning allows for updating deep models with several new samples, which may provide incremental demands from users in wireless sensing model services. For example, OneFi\cite{onefi} stands as a representative work of few-shot learning, with its core concept being the ability to recognize unseen gestures using only one (or a few) labeled samples. This method, after learning 20 base classes in the pre-training stage, can rapidly learn new classes never seen before with a small amount of data samples through fine-tuning. OneFi primarily emphasizes how to quickly adapt to new classes, but it can not recognize old action classes. In contrast, our CCS effectively preserves and updates the knowledge of old action classes by introducing a memory mechanism during model training, thus it is suitable for continuously incremental model services.
Another work, Wi-Fringe\cite{wifringe}, is a device-free human activity recognition system based on WiFi data. Leveraging the principle of zero-shot learning, it combines radio frequency signal features with attribute and context-aware representations of English words or phrases, enabling inference of activities without prior training samples. While Wi-Fringe can identify some unseen activity categories, this capability relies on the similarity of label texts and cannot recognize new categories that lack semantic relevance with previous categories in text. This implies a limitation of Wi-Fringe when dealing with new categories significantly different from previous ones in terms of semantics, as it cannot utilize existing text information to infer activities of these new categories. In contrast, our proposed CCS does not rely on text information. Utilizing the method of class-incremental learning, it addresses catastrophic forgetting while accommodating the learning of new category demands. This enables CCS to better meet the needs of different users without being restricted by the similarity of text information, while maintaining good generalization capability.
\subsection{Continuous Learning}
Continuous learning or incremental learning aims for ``learning without forgetting''~\cite{lwf}, which can be utilized to provide incremental demands from users in wireless sensing model services. Regularization\cite{regular1,regular2,regular3,regular4} and data replay\cite{replay1,regular1,replay3} are two common incremental learning methods in human action recognition tasks. One representative regularization method is knowledge distillation\cite{lwf,kd1,kd2}. Its core concept involves introducing an additional loss function to guide the training of the student model using the knowledge from the teacher model, facilitating better learning and generalization. The goal of this process is to enable the student model to mimic the behavior of the teacher model as closely as possible. It not only employs the loss function of the true labels but also introduces an extra loss function to measure the similarity between the predictions of the student and teacher models, often using cross-entropy loss or other forms of distance metrics. Since the regularization method does not store existing activity samples, its recognition accuracy for existing activities is weak. iCaRL~\cite{icarl} represents the data replay method and, in addition to employing knowledge distillation, also selects a subset of samples from known classes in each incremental stage and incorporates them into the exemplar set along with the data from newly added classes for training. For the selection of exemplar samples, random selection is the most primitive method, and Herding~\cite{Herding} has become a commonly used strategy for selecting exemplars. In this paper, we also adopt the Herding strategy for selecting exemplar samples.
\section{Methodology}\label{sec:methodology}
\subsection{Problem State}\label{sec:problem_state}
As shown in Fig.~\ref{fig:layout}, CCS enables wireless sensing service providers provide base model service, incremental model service, and continuously incremental model service for users. We term these service stages as $S_i, i\in [0, n]$, where $S_0$ represents the base model service; $S_i$($i>0$) represents the incremental service at the $i$-th stage. In $S_0$, the service provider trains a base model $\mathcal{M}_0$ with the dataset from the provider $D^p$, and distribute to users' local devices as the base model service.
\begin{figure*}[t]
\centering
\includegraphics[width=0.8\textwidth]{figs/fig_2.pdf}
\caption{We incorporate knowledge distillation and weight aligning throughout the training process of each incremental stage. During each incremental stage, exemplars are extracted from the training data of the previous stage using the Herding method~\cite{Herding}. The network learns to retain knowledge of old tasks by employing knowledge distillation from a frozen model of the previous stage, utilizing MSE loss for distillation. Additionally, weight aligning addresses the issue of imbalance in the weight distribution between new and old services.}
\label{fig:layout}
\end{figure*}
As the first incremental model services, a user generate new demands. To align $\mathcal{M}_0$ with these needs, CCS necessitates that users locally prepare data, $D_1^u$, relevant to the new demand for updating $\mathcal{M}_0$ . If $\mathcal{M}_0$ were solely updated with $D_1^u$, it would likely suffer from catastrophic forgetting, losing its initial capabilities. Addressing this issue, CCS requires the service provider to select a subset of data (exemplars) $E^p$ from $D^p$, to be transmitted to the user's local device for inclusion in the model update process, yielding $\mathcal{M}_1$. The objective of CCS is to ensure that $\mathcal{M}_1$ not only accommodates the user's new requirements but also retains its original capabilities. This process can be represented as Equation.~\ref{eq:m1}.
\begin{equation}\label{eq:m1}
\mathcal{M}_1 = \text{CCS}(\mathcal{M}_0, D_1^u, E^p)
\end{equation}
Following this logic, when updating $\mathcal{M}_i$($i\ge2$), the process can be represented as Equation.~\ref{eq:ms}.
\begin{equation}\label{eq:ms}
\mathcal{M}_i = \text{CCS}(\mathcal{M}_{i-1}, D_i^u, E^p, E_1^u, E_2^u,...,E_{i-1}^u), i\ge2
\end{equation}
where $D_i^u$ represents user's dataset for new demands in stage $S_{i}$; $E_{i-1}^u$ represents exemplars from $D_{i-1}^u$.
Next, we will introduce how to select exemplars in Sec.~\ref{sec:exemplar} and how to update models in Sec.~\ref{sec:knowledge} and Sec.~\ref{sec:weight} in detail.
\subsection{Exemplar Selection}\label{sec:exemplar}
% The selection and update methods of these exemplar sets directly influence the performance and accuracy of the model, especially in adapting to evolving data environments.}
To mitigate catastrophic forgetting in continuous learning, selecting exemplars in previous learning stage has been widely demonstrated as an effective method~\cite{rainbow,gradient,anchor}. For example, one can choose the centroids of K-Means clustering as exemplars. In this paper, we apply the more advanced Herding method~\cite{Herding} to select exemplars. Suppose the provider's data $D^p$ includes $C$ classes of data for services, for $c$-th class, its data sample is denoted as $d_i^c, i\in[1,N_c]$, where $N_c$ represents the total number of data samples of the $c$-th class. We compute the normalized features of $d_i^c, i\in[1,N_c]$, and take the average value as the feature center for the $c$-th class of the provider's data.
\begin{equation}\label{eq:featurecenter}
F_c^p = \frac{1}{N_c} \times \sum_{i=1}^{N_c} \text{Norm}(\mathcal{M}_0(d_i^c))
\end{equation}
We select the $K$ data samples that are nearest (Euclidean distance) to this feature center as exemplar subset $E_c^p$.
% \textcolor{red}{The Equation.~\ref{eq:center_distance} illustrates the method for calculating the distance between samples and the class center. In this paper, the typicality of the samples is determined by calculating the Euclidean distance between the class centers and the feature layers.}
By repeating the above process for all classes within the provider's data $D^p$, we can acquire the exemplars $E^p$. Therefore, $E^p$ includes $C\times K$ samples. A smaller K indicates that the selected exemplars are more compact or sparse. The exemplars $E^p$ then participates in updating $\mathcal{M}_1$ in Equation.~\ref{eq:m1} for the incremental model service.
% \begin{equation}\label{eq:center_distance}
% D(F_c^p, d_i^c) = \left\| \sum_{j=1}^{K} (F_c^p - \mathcal{M}_0(d_i^c)) \right\|
% \end{equation}
Following this logic, we can select user's exemplars $E_{i-1}^u$ at stage $S_i$ for continuously incremental model service, shown in Equation.~\ref{eq:ms}. The upper part of Fig.~\ref{fig:layout} illustrates how the number of exemplars involved in model updates accumulates over each stage.
\subsection{Knowledge Distillation}\label{sec:knowledge}
Knowledge distillation involves training a student model $\mathcal{M}_i$ to mimic and learn from a teacher model $\mathcal{M}_{i-1}$, which has already acquired knowledge of old knowledge. This process not only mitigates catastrophic forgetting during the learning of new knowledge but also effectively leverages previously acquired knowledge while maintaining overall model performance. This section details how the model $\mathcal{M}_i$ is updated in knowledge distillation, utilizing the user data corresponding to the new demands of the current stage ($D_i^u$), the model from the previous stage $\mathcal{M}_{i-1}$, and the set of exemplars. The ultimate objective of the method is to ensure that $\mathcal{M}_i$, while meeting the new demands, also ideally retains the capabilities of serving the older functions.
As illustrated in Fig.~\ref{fig:layout}, CCS initially sets $\mathcal{M}_{i-1}$ as the starting point for $\mathcal{M}_i$ and freezes $\mathcal{M}_{i-1}$ to act as the teacher network for $\mathcal{M}_i$. It is noteworthy that, during the $i$-th stage, new services are added; consequently, prediction head of $\mathcal{M}_{i-1}$ expands to accommodate services for the new task. The loss function to updates the model is as follows.
\begin{equation}\label{eq:updateloss}
Loss = (1-\alpha) \text{Dis}(\mathcal{M}_i(x), y) + \alpha \text{MSE}( \mathcal{M}_i(x), \mathcal{M}_{i-1}(x))
\end{equation}
where $(x,y)$ represents a pair consisting of a training sample and its corresponding label. $\text{Dis}(\mathcal{M}_i(x), y)$ represents the distance between the prediction of $\mathcal{M}_s$ and the true value, responsible for enabling the model to acquire new service capabilities. $\text{MSE}( \mathcal{M}_i(x), \mathcal{M}_{i-1}(x))$ represents the MSE loss between output of $\mathcal{M}_i$ and $\mathcal{M}_{i-1}$, distilling knowledge from $\mathcal{M}_{i-1}$ to $\mathcal{M}_i$. $\alpha$ is to balance the two distances.
\subsection{Weight Aligning} \label{sec:weight}
Introducing new service tasks can disrupt model eight allocation, causing model imbalance and reducing learning capacity. This can lead to suboptimal performance on new tasks and catastrophic forgetting of previously learned tasks. In response to these challenges, we introduces additional weight adjustment mechanisms during model training to effectively balance the distribution of weights between new and old tasks, described next.
Assuming $M_{i-1}$ is capable of providing $u$ old services, and $M_{i}$ introduces $v$ new services, the weights of prediction head of $M_{i}$ can accordingly be represented as Equation.~\ref{eq:weight1}.
\begin{equation}\label{eq:weight1}
W^{i} = (w_1,w_2,...,w_{u},w_{u+1},...,w_{u+v} ).
\end{equation}
We leverage weight aligning technique~\cite{wa} to re-balance the weights between old services and new services as Equation.~\ref{eq:weight_align}.
\begin{equation}\label{eq:weight_align}
\begin{gathered}
\text{Norm}_{old} = (\left\|w_1\right\|,\left\|w_2\right\|,...,\left\|{w_u}\right\|)\\
\text{Norm}_{new} = (\left\|w_{u+1}\right\|,...,\left\|w_{u+v}\right\|)\\
(\widehat{w}_{u+1},...,\widehat{w}_{u+v}) = \frac{ \text{Mean} (\text{Norm}_{old} )}{ \text{Mean}(\text{Norm}_{new} )} \times (w_{u+1},...,w_{u+v} )
\end{gathered}
\end{equation}
where $\widehat{w}_{u+i},i\in[1,v]$ represent the weights for new services after weight aligning. After weight aligning, the weights of prediction head of model $M_{i}$ turns to be
\begin{equation}\label{eq:weight_model-i}
\widehat{W}^{i} = (w_1,w_2,...,w_{u}, \widehat{w}_{u+1},...,\widehat{w}_{u+v} ).
\end{equation}
\subsection{Hyper-parameter Setting}
For each new service class added, we select one exemplar, thus setting K = 1, mentioned in Sec.~\ref{sec:exemplar}. We set $ \alpha = 0.1\times u/(u+v)$ in Equation.~\ref{eq:updateloss} to balance the two distances for model updating. We set $\text{Dis}()$ as cross-entropy loss for the classification task.
\section{Experiments}\label{sec:experiments}
\subsection{Evaluation Dataset}\label{sec:dataset}
We use the large-scale XRF55 dataset~\cite{xrf55} to evaluate CCS, primarily for two reasons: (1) The XRF55 dataset encompasses three distinct wireless modalities – Wi-Fi, millimeter-wave radar, and RFID – allowing us to test CCS's adaptability and versatility across different wireless modalities; (2) Comprising 55 categories of action data, XRF55 boasts the broadest range of action classes currently available, facilitating the continuous addition of new classification tasks as users' needs, which in turn enables a comprehensive assessment of CCS's capability for incremental model services and continuously incremental model service. We experimentally employ ResNets~\cite{he2016deep} as the sevice models for RFID and Wi-Fi modalites, and empirically employ Temporal UNet~\cite{wang2019temporal} as the service model for mmWave radar modality.
% \begin{table}[ht!]
\begin{table}[t]
\centering
\setlength{\tabcolsep}{2pt} % 调整列间距
\resizebox{\columnwidth}{!}{%
\begin{tabular}{ccccc}
\hline
Action & Person & Modality & Training samples & Test samples \\ \hline
55 & 30 & Wi-Fi, mmWave, RFID & 23100 & 9900 \\ \hline
\end{tabular}
}
\caption{The XRF55 dataset~\cite{xrf55} encompasses 55 distinct actions across three wireless modalities. We use its samples at the scene \#1, totaling 23100 training samples and 9900 testing samples.}
\label{tab:xrf55}
\end{table}
\begin{table}[t]
\centering
\setlength{\tabcolsep}{2.5pt} % 调整列间距
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccccccc}
\toprule
& Stage 1 & Stage 2 & Stage 3 & Stage 4 & Stage 5 & Result table & Result figure \\
\midrule
User1 & C1$\sim$C15 & C16$\sim$C25 & C26$\sim$C35 & C36$\sim$C45 & C46$\sim$C55 & Table.~\ref{tab:result1} & Fig.~\ref{fig:result1} \\
User2 & C1$\sim$C15 & C26$\sim$C35 & C46$\sim$C55 & C36$\sim$C45 & C16$\sim$C25 & Table.~\ref{tab:result2} & Fig.~\ref{fig:result2} \\
User3 & C1$\sim$C15 & C36$\sim$C45 & C46$\sim$C55 & C16$\sim$C25 & C26$\sim$C35 & Table.~\ref{tab:result3} & Fig.~\ref{fig:result3} \\
User4 & C1$\sim$C15 & C46$\sim$C55 & C16$\sim$C25 & C26$\sim$C35 & C36$\sim$C45 & Table.~\ref{tab:result4} & Fig.~\ref{fig:result4}\\
\bottomrule
\end{tabular}
% }
}
\caption{ We assume four users require customized incremental services in differing sequences. We assess CCS under this configuration. The results are shown in corresponding tables and figures as listed.}
\label{tab:user_sequences}
\end{table}
To simulate the base model service in CCS and the scenario where users continuously add new requirements, we randomly divided the XRF55 dataset into five subsets based on action categories, with each subset containing 15, 10, 10, 10, and 10 classes of action data, respectively. The subset with 15 classes is utilized to train the base model in CCS, while the remaining subsets are sequentially introduced as new action recognition demands, serving to assess the capability of CCS's incremental model services and continuously incremental model service. For simplicity, we denote the first subset, which includes action categories 1 through 15, as C1$\sim$C15. The second subset, comprising categories 16 to 25, is denoted C16$\sim$C25. Following this pattern, the remaining three subsets, encompassing subsequent action categories, are correspondingly named as C26$\sim$C35, C36$\sim$C45, and C46$\sim$C55, respectively.
Table.~\ref{tab:user_sequences} illustrates our evaluation scheme, in which we assume four users requiring customized incremental services in differing sequences. Samples of C1$\sim$C15 serve as the base model service in the first stage, and the remaining subsets representing incremental services. Under this configuration, we assess the performance of CCS, and report the results in upcoming figures and tables of this section, specified in Table.~\ref{tab:user_sequences}.
In stages 2-5, we employ the Herding~\cite{Herding} method to select exemplars from the preceding stage, at a ratio of one sample per person per action, for model updating. For instance, in stage 2, we have $1\times30\times 15=450$ exemplars to update models, which is $1/14\approx 7\%$ of training data in stage 1.
% XRFDataset contains three modalities: RFID, Wi-Fi and millimeter wave. All modalities consist of samples from fifty-five action classes. XRFDataset provides prepoccessed data of all three modalities, where Wi-Fi$\in 1000\times270$, RFID$\in148\times23$, mmWave $\in17\times256\times128$.
% We utilize provided RFID and Wi-Fi data to conduct the following experiments. Assume mmWave preprocessed data provided from XRFDataset has dimension of T × R × V. We sum values on dimension V for any R and any T to further reduce dimension of data. Therefore, the dimension of mmWave data used for the following experiments is $\in17\times256$
% Assume we have one initialization phase and four incremental phases. The fifty-five action classes are randomly shuffled and divided into five mutually disjoint sub dataset. One contains fifteen action classes which is the base set for initialization phase. The other four sub datasets contain ten action classes for each and are used for different incremental stage. For each class, exemplar can only store 30 typical samples.
\subsection{Evaluation Metrics}
\textbf{(1) Accuracy } We use classification accuracy to indicate CCS's ability to classify actions at different service stages.
% The equation for computing accuracy is denoted as Equation.~\ref{eq:accuracy}.
% \begin{equation}\label{eq:accuracy}
% Accuracy= \frac{1}{N} \sum_{i=1}^{N} I(y_{i}, pre_{i})
% \end{equation}
% where $pre_{i}$, $y_{i}$ represents the prediction and label of the $i^{th}$ sample. $I(y_{i}, pre_{i})$ is an indicator which outputs 1 if the prediction and the label are the same and output 0 otherwise.
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{User1} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C16$\sim$C25} & \textbf{+C26$\sim$C35} & \textbf{+C36$\sim$C45} & \textbf{+C46$\sim$C55} \\
\midrule
\multirow{5}{*}{RFID} & Baseline & 75.96 & 32.82 & 20.24 & 17.42 & 13.43 \\
& iCaRL & 75.85 & 48.87 & 33.29 & 30.54 & 26.67 \\
& UCIR & 74.30 & 49.38 & 34.51 & 31.94 & 27.71 \\
& BiC & 73.81 & 55.02 & 36.56 & 24.31 & 23.55 \\
& OneFi & 58.81 & 14.44 & 7.21 & 5.75 & 6.47 \\
& Ours & 75.96 & 58.56 & 44.89 & 38.23 & 34.01 \\
\midrule
\multirow{5}{*}{Wi-Fi} & Baseline & 95.07 & 39.09 & 26.54 & 20.86 & 18.33 \\
& iCaRL & 94.85 & 72.51 & 65.86 & 56.01 & 56.57 \\
& UCIR & 94.89 & 79.4 & 71.37 & 58.85 & 60.2 \\
& BiC & 93.37 & 74.13 & 58.63 & 52.58 & 53.67 \\
& OneFi & 84.30 & 20.24 & 10.40 & 10.32 & 8.64 \\
& Ours & 95.07 & 83.56 & 75.48 & 64.12 & 65.10 \\
\midrule
\multirow{5}{*}{mmWave} & Baseline & 91.11 & 38.29 & 25.24 & 20.47 & 17.47 \\
& iCaRL & 91.70 & 77.58 & 68.33 & 58.48 & 56.58 \\
& UCIR & 91.41 & 80.60 & 72.38 & 64.48 & 63.45 \\
& BiC & 91.30 & 79.29 & 73.33 & 66.59 & 59.99 \\
& OneFi & 84.37 & 31.87 & 15.13 & 12.27 & 11.97 \\
& Ours & 93.07 & 87.49 & 81.16 & 69.35 & 69.19 \\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{
% \textcolor{blue}{
Accuracy of CCS services for user1 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS outperforms existing alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi} by a large margin for three wireless modalities.
% }
}
\label{tab:result1}
\end{table}
\begin{figure}[t]
\centering
% \subfloat[Accuracy]{\includesvg[width=\linewidth]{figs/table1.svg}} \\
\includegraphics[width=\linewidth]{figs/accn1.pdf}
\caption{
% \textcolor{blue}{
ACCN of CCS services for user1 with continuous stages listed in Table.~\ref{tab:user_sequences}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi}, demonstrating a consistent rise in model value.
% }
}
\label{fig:result1}
\end{figure}
\textbf{(2) ACCN } CCS's mode capacity, the number of recognizable action categories, would increase during incremental service, however, the average accuracy might decline during the forgetting old actions in incremental learning. We adopt a novel metric to assess model worth – the multiplication of the number of recognizable action categories and model accuracy~(ACCN), which can balance the model capacity and accuracy.
\begin{equation}\label{eq:accn}
ACCN = N \times Accuracy
\end{equation}
where $N$ is the number of recognizable action categories by CCS models.
% \subsection{Comparison Study}
\subsection{Results}\label{sec:results}
% In Sec.~\ref{sec:dataset}, we outlined the division of the XRF55 into five subsets, with C1$\sim$C15 serving as the base model service in the first stage, and the remaining subsets representing incremental services. Table.~\ref{tab:user_sequences} illustrates our evaluation scheme, in which we assume four hypothetical users requiring customized incremental services in differing sequences. Under this configuration, we assess the performance of CCS, and report the results in upcoming figures and tables of this section, specified in Table.~\ref{tab:user_sequences}. person. In stages 2-5, we employ the Herding~\cite{Herding} method to select exemplars from the preceding stage, at a ratio of one sample per person per action, for model updating. For instance, in stage 2, we have $1\times30\times 15=450$ exemplars to update base models in stage 1.
% \begin{table}[htb]
% \resizebox{\columnwidth}{!}{%
% \begin{tabular}{cccccccc}
% \toprule
% & Stage 1 & Stage 2 & Stage 3 & Stage 4 & Stage 5 & Result table & Result figure \\
% \midrule
% User1 & C1$\sim$C15 & C16$\sim$C25 & C26$\sim$C35 & C36$\sim$C45 & C46$\sim$C55 & Table.~\ref{tab:result1} & Fig.~\ref{fig:result1} \\
% User2 & C1$\sim$C15 & C26$\sim$C35 & C46$\sim$C55 & C36$\sim$C45 & C16$\sim$C25 & Table.~\ref{tab:result2} & Fig.~\ref{fig:result2} \\
% User3 & C1$\sim$C15 & C36$\sim$C45 & C46$\sim$C55 & C16$\sim$C25 & C26$\sim$C35 & Table.~\ref{tab:result3} & Fig.~\ref{fig:result3} \\
% User4 & C1$\sim$C15 & C46$\sim$C55 & C16$\sim$C25 & C26$\sim$C35 & C36$\sim$C45 & Table.~\ref{tab:result4} & Fig .~\ref{fig:result4}\\
% \bottomrule
% \end{tabular}
% }
% \caption{ We assume four hypothetical users requiring customized incremental services in differing sequences. We assess the performance of CCS under this configuration. }
% \label{tab:user_sequences}
% \vspace{-1.5em}
% \end{table}
% User1 increments according to the sequence C1$\sim$C15 C16$\sim$C25 C26$\sim$C35 C36$\sim$C45 C46$\sim$C55, with accuracy results displayed in Table.~\ref{tab:result1}, and the ACCN curve shown in Fig.~\ref{fig:result1}. User2 increments according to the sequence C1$\sim$C15 C26$\sim$C35 C46$\sim$C55 C36$\sim$C45 C16$\sim$C25, with accuracy results displayed in Table.~\ref{tab:result2}, and the ACCN curve shown in Fig .~\ref{fig:result2}. User3 increments according to the sequence C1$\sim$C15 C36$\sim$C45 C46$\sim$C55 C16$\sim$C25 C26$\sim$C35, with accuracy results displayed in Table.~\ref{tab:result3}, and the ACCN curve shown in Fig .~\ref{fig:result3}. User4 increments according to the sequence C1$\sim$C15 C46$\sim$C55 C16$\sim$C25 C26$\sim$C35 C36$\sim$C45, with accuracy results displayed in Table.~\ref{tab:result4}, and the ACCN curve shown in Fig .~\ref{fig:result4}.
% We compare our method with the typical methods for class incremental learning, which include iCaRL, UCIR and BiC.
% The baseline method here calculates action recognition accuracies for all seen and unseen action classes at different stage using cross-entropy loss function, and it does not include exemplar data during the training process at each stage. The base classes introduced in initialization stage are denoted as $C_1\sim C_{15}$, and unseen classes introduced in incremental stages are $C_{16}\sim C_{25}$, $C_{26} \sim C_{35}$, $ C_{36} \sim C_{45}$, $C_{46} \sim C_{55}$ respectively. Tables xx present the overall action recognition accuracies of all three modality at different stage. The results indicate that our method has better performance in all three modalities relatively.
\textbf{(1) Accuracy and ACCN on User1}
In Table.~\ref{tab:result1}, we show CCS's accuracy on user1 at five continuous stages listed in Table.~\ref{tab:user_sequences}. The baseline accuracy is derived from updating models with only new data in the new stage, which cannot recognize old actions even starting at the second stage. It is observable that CCS exhibits high accuracy in the first stage, particularly with Wi-Fi and RFID data. As the number of stages increases, the accuracy gradually declines. By the final stage, the respective accuracy for RFID, Wi-Fi, and millimeter-wave radar are 34.01\%, 65.10\%, and 69.19\%, respectively. Nonetheless, it remains notably higher than current alternative methods like iCaRL~\cite{icarl}, UCIR~\cite{ucir}, and BiC~\cite{bic} by a large margin. CCS also largely outperforms OneFi~\cite{onefi}, which is proposed to recognize unseen actions with one/few wireless samples. OneFi focus on enabling the model to quickly transfer and learn new action knowledge, without addressing the issue of catastrophic forgetting. As a result, OneFi almost completely loses its ability to recognize old actions during the incremental learning stage.
% \textcolor{red}{Additionally, we also reproduced the OneFi method mentioned in Sec.~\ref{sec:related_work}, which is used for recognizing human activities through wireless signals.
% In the experiments, the OneFi method uses the same test set and initial training set as the CCS setup. However, in the incremental stages, the training set used by OneFi can be considered a subset of the training set used by CCS. Specifically, in the incremental stages of the OneFi method, each new category contains only thirty data samples. Due to the significantly fewer training samples used by OneFi at each stage compared to other methods, the accuracy of the model in recognizing new categories is also suboptimal.
% The OneFi method performs poorly in balancing new and old categories, leading to a significant loss of memory for old categories during incremental learning.
% Additionally, due to the limited number of training samples, the accuracy of the OneFi method in recognizing new categories is also low.
% In contrast, the CCS method shows a more consistent performance improvement, closely aligning with the ideal curve, demonstrating its effectiveness in handling incremental learning tasks. These results clearly indicate that the CCS method significantly outperforms the OneFi method in incremental learning, both in mitigating catastrophic forgetting and improving the recognition accuracy of new categories.
% }
Fig.~\ref{fig:result1} shows ACCN curves, where the ideal scenario would be one where the model learns new information without forgetting old information; hence, the ACCN curve would linearly increase with the incremental capacity of the model. In reality, however, models partially forget previously learned knowledge. As depicted, all methods fall short of this ideal, yet, CCS, traces the closest path to this ideal curve, demonstrating a consistent rise in model value measured by ACCN.
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{User2} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C26$\sim$C35} & \textbf{+C46$\sim$C55} & \textbf{+C36$\sim$C45} & \textbf{+C16$\sim$C25} \\
\midrule
\multirow{5}{*}{RFID} & Baseline & 72.96 & 27.40 & 20.83 & 16.81 & 14.58 \\
& iCaRL & 74.15 & 43.84 & 29.89 & 26.98 & 21.43 \\
& UCIR & 73.70 & 42.24 & 30.84 & 29.93 & 25.03 \\
& BiC & 73.81 & 37.96 & 31.73 & 25.07 & 23.84 \\
& OneFi & 58.81 & 10.87 & 10.05 & 5.74 & 5.74 \\
& Ours & 74.30 & 52.31 & 39.83 & 34.64 & 32.14 \\
\midrule
\multirow{5}{*}{Wi-Fi} & Baseline & 95.07 & 37.02 & 28.16 & 20.46 & 17.77 \\
& iCaRL & 94.85 & 72.71 & 63.24 & 54.69 & 54.1\\
& UCIR & 94.89 & 80.16 & 71.22 & 57.26 & 59.91 \\
& BiC & 93.37 & 72.09 & 60.87 & 57.89 & 55.23 \\
& OneFi & 84.30 & 17.31 & 13.98 & 10.33 & 8.99 \\
& Ours & 95.07 & 83.62 & 76.00 & 63.67 & 64.36 \\
\midrule
\multirow{5}{*}{mmWave} & Baseline & 91.11 & 35.2 & 27.49 & 20.47 & 17.32 \\
& iCaRL & 91.70 & 73.22 & 64.6 & 55.7 & 55.82\\
& UCIR & 91.41 & 76.36 & 72.0 & 62.21 & 61.79\\
& BiC & 91.30 & 77.0 & 70.49 & 64.89 & 63.83\\
& OneFi & 84.37 & 35.38 & 21.56 & 13.51 & 12.74 \\
& Ours & 93.07 & 84.11 & 78.7 & 68.96 & 70.77\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Accuracy of CCS services for user2 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS outperforms existing alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi} by a large margin for three wireless modalities. }
\label{tab:result2}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/accn2.pdf}
\caption{ ACCN of CCS services for user2 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi}, demonstrating a consistent rise in model value. }
\label{fig:result2}
\end{figure}
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{User3} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C36$\sim$C45} & \textbf{+C46$\sim$C55} & \textbf{+C16$\sim$C25} & \textbf{+C26$\sim$C35} \\
\midrule
\multirow{5}{*}{RFID} & Baseline & 74.04 & 32.02 & 21.43 & 18.46 & 12.07\\
& iCaRL & 73.74 & 41.47 & 30.03 & 28.37 & 23.28\\
& UCIR & 73.70 & 49.53 & 34.76 & 31.19 & 24.64\\
& BiC & 73.81 & 37.67 & 32.60 & 27.90 & 23.36\\
& OneFi & 58.81 & 13.09 & 10.19 & 7.23 & 4.34 \\
& Ours & 74.30 & 58.29 & 45.21 & 40.23 & 33.86\\
\midrule
\multirow{5}{*}{Wi-Fi} & Baseline & 95.07 & 37.38 & 28.3 & 21.85 & 17.14\\
& iCaRL & 94.85 & 73.2 & 65.78 & 58.83 & 57.13\\
& UCIR & 94.89 & 80.44 & 71.19 & 62.44 & 61.12\\
& BiC & 93.37 & 73.71 & 62.95 & 60.12 & 54.03\\
& OneFi & 84.30 & 18.78 & 13.95 & 11.04 & 6.79 \\
& Ours & 95.07 & 83.4 & 78.38 & 66.72 & 63.62\\
\midrule
\multirow{5}{*}{mmWave} & Baseline & 91.11 & 36.8 & 27.43 & 21.35 & 16.16\\
& iCaRL & 91.7 & 72.4 & 64.76 & 61.35 & 57.96\\
& UCIR & 91.41 & 75.36 & 69.87 & 67.81 & 63.27\\
& BiC & 91.3 & 75.02 & 70.22 & 69.27 & 65.24\\
& OneFi & 84.37 & 28.24 & 21.76 & 15.56 & 9.05 \\
& Ours & 93.07 & 84.64 & 78.41 & 71.70 & 69.78\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Accuracy of CCS services for user3 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS outperforms existing alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi} by a large margin for three wireless modalities. }
\label{tab:result3}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/accn3.pdf}
\caption{ ACCN of CCS services for user3 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi}, demonstrating a consistent rise in model value. }
\label{fig:result3}
\end{figure}
\textbf{(2) Accuracy and ACCN on User2}
In Table.~\ref{tab:result2}, we show CCS's accuracy on user2 at five continuous stages listed in Table.~\ref{tab:user_sequences}. Although CCS follows a different stage sequence for user2 compared to user1, which progresses incrementally in the order of C1$\sim$15, C26$\sim$C35, C46$\sim$C55, C36$\sim$C45, C16$\sim$C25, it consistently demonstrates high accuracy across all stages. By the end of the final stage, the accuracy rates for RFID, Wi-Fi, and millimeter-wave radar are 32.56\%, 63.87\%, and 67.45\%, correspondingly. Fig.~\ref{fig:result2} illustrates the ACCN curve of CCS services for user2. Our proposed method, CCS, traces the closest path to this ideal curve compared to current alternative methods like iCaRL~\cite{icarl}, UCIR~\cite{ucir}, and BiC~\cite{bic}, indicating a consistent rise in model value, measured by ACCN.
% }
% {\color{red}
\textbf{(3) Accuracy and ACCN on User3}
In Table.~\ref{tab:result3}, we show CCS's accuracy on user3 at five continuous stages listed in Table.~\ref{tab:user_sequences}. Our method significantly mitigates models’ accuracy degradation. In final stage, our method achieves accuracy improvements of 9.22\%, 2.5\%, and 4.54\% in RFID, Wi-Fi and mmWave modalities respectively, compared to the second best method. Fig.~\ref{fig:result3} illustrates the ACCN curve of CCS services for user3. Our proposed solution can reap highest model value across all stages and modalites.
% }
\textbf{(4) Accuracy and ACCN on User4}
In Table.~\ref{tab:result4}, we show CCS's accuracy on user4 at five continuous stages listed in Table.~\ref{tab:user_sequences}. Our method significantly mitigates models’ accuracy degradation. In the final stage, our method achieves accuracy improvements of 5.92\%,2.16\% and 6.71\% in RFID, Wi-Fi, and mmWave modalities, respectively, compared to the second best method. Fig.~\ref{fig:result4} illustrates the ACCN curve of CCS services for user4. Our proposed solution has the potential to yield the highest model value across all stages and modalities.
The results in Sec.~\ref{sec:results} show that CCS consistently performs best across different incremental demand sequences for four users. Additionally, it demonstrates increasing model value over time, highlighting CCS's versatility in addressing customized incremental wireless sensing services.
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{User4} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C46$\sim$C55} & \textbf{+C16$\sim$C25} & \textbf{+C26$\sim$C35} & \textbf{+C36$\sim$C45} \\
\midrule
\multirow{5}{*}{RFID} & Baseline & 74.04 & 28.91 & 22.65 & 14.63 & 14.07\\
& iCaRL & 73.74 & 36.93 & 27.25 & 27.27 & 23.59\\
& UCIR & 73.70 & 44.20 & 31.63 & 34.09 & 27.23\\
& BiC & 73.81 & 40.18 & 36.19 & 27.47 & 20.36\\
& OneFi & 58.81 & 14.31 & 9.30 & 5.51 & 4.53 \\
& Ours & 74.30 & 52.07 & 42.21 & 37.17 & 33.15\\
\midrule
\multirow{5}{*}{Wi-Fi} & Baseline & 95.07 & 39.53 & 27.98 & 20.74 & 17.12\\
& iCaRL & 94.85 & 72.96 & 64.68 & 57.62 & 57.81\\
& UCIR & 94.89 & 82.18 & 70.9 & 58.23 & 60.01\\
& BiC & 93.37 & 78.11 & 63.46 & 46.91 & 43.55\\
& OneFi & 84.30 & 23.84 & 14.70 & 8.51 & 8.28 \\
& Ours & 95.07 & 84.82 & 77.54 & 62.19 & 62.17\\
\midrule
\multirow{5}{*}{mmWave} & Baseline & 91.11 & 38.60 & 27.35 & 19.60 & 16.86\\
& iCaRL & 91.70 & 74.38 & 71.02 & 62.10 & 58.91\\
& UCIR & 91.41 & 76.96 & 75.17 & 67.31 & 63.66\\
& BiC & 91.30 & 79.22 & 73.06 & 67.38 & 60.16\\
& OneFi & 84.37 & 31.69 & 19.94 & 11.37 & 10.02 \\
& Ours & 93.07 & 84.00 & 82.59 & 72.98 & 70.37\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Accuracy of CCS services for user4 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS outperforms existing alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi} by a large margin for three wireless modalities. }
\label{tab:result4}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/accn4.pdf}
\caption{ ACCN of CCS services for user4 with the continuous demands listed in Table.~\ref{tab:user_sequences}. CCS traces the closest path to the ideal curve compared to alternative methods such as iCaRL~\cite{icarl}, UCIR~\cite{ucir}, BiC~\cite{bic} and OneFi~\cite{onefi}, demonstrating a consistent rise in model value. }
\label{fig:result4}
\end{figure}
% {\color{red}
\subsection{Ablation Study}\label{sec:ablation}
In this section, we first conduct ablation study on the effectiveness of the combination among the applied techniques in CCS, including exemplar selection, knowledge distillation and weight aligning. Besides, we evaluate different loss functions in knowledge distillation and different normalization method in weight aligning.
% {
% \color{red}
\begin{table}[t]
\small
\centering
\setlength{\tabcolsep}{2pt} % 调整列间距
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccc}
\toprule
\multirow{2}{*}{\textbf{Modality}} & \multicolumn{3}{c}{\textbf{Method}} & \multicolumn{1}{c}{\multirow{2}{*}{C1$\sim$C15}} & \multicolumn{1}{c}{\multirow{2}{*}{+C16$\sim$C25}} & \multicolumn{1}{c}{\multirow{2}{*}{+C26$\sim$C35}} & \multicolumn{1}{c}{\multirow{2}{*}{+C36$\sim$C45}} & \multicolumn{1}{c}{\multirow{2}{*}{+C46$\sim$C55}} \\
& \multicolumn{1}{c}{\textbf{E}} & \multicolumn{1}{c}{\textbf{KD}} & \multicolumn{1}{c}{\textbf{WA}} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{}\\
\midrule
\multirow{5}{*}{RFID}
& / & \checkmark & \checkmark & 74.04 & 33.24 & 23.06 & 19.41 & 14.72\\
& \checkmark & / & / & 74.04 & 43.53 & 28.33 & 28.07 & 20.13\\
& \checkmark & \checkmark & / & 74.30 & 48.20 & 33.75 & 28.59 & 22.87\\
& \checkmark & / & \checkmark & 74.30 & 48.80 & 35.41 & 37.27 & 30.06\\
& \checkmark & \checkmark & \checkmark & 75.96 & 58.56 & 44.89 & 38.23 & 34.01\\
\midrule
\multirow{5}{*}{Wi-Fi}
& / & \checkmark & \checkmark & 95.07 & 41.53 & 33.76 & 25.98 & 24.20\\
& \checkmark & / & / & 95.07 & 72.00 & 65.57 & 52.22 & 56.01\\
& \checkmark & \checkmark & / & 94.85 & 73.38 & 66.89 & 55.98 & 56.87\\
& \checkmark & / & \checkmark & 95.07 & 76.18 & 69.40 & 56.95 & 61.08\\
& \checkmark & \checkmark & \checkmark & 95.07 & 83.56 & 75.48 & 64.12 & 65.10\\
\midrule
\multirow{5}{*}{mmWave}
& / & \checkmark & \checkmark & 91.11 & 62.38 & 47.86 & 41.20 & 35.29\\
& \checkmark & / & / & 93.07 & 76.38 & 66.21 & 57.96 & 56.14\\
& \checkmark & \checkmark & / & 93.07 & 84.07 & 73.97 & 62.74 & 59.45\\
& \checkmark & / & \checkmark & 93.07 & 80.91 & 73.24 & 60.72 & 58.25\\
& \checkmark & \checkmark & \checkmark & 93.07 & 87.49 & 81.16 & 69.35 & 69.19\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Accuracy of CCS services in ablation study for user1 with the continuous stages listed in Table.\ref{tab:user_sequences}. E, KD, WA represent Herding exemplar selection, knowledge distillation and weight aligning respectively.}
\label{tab:ablation_ekdwa}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/ab_ekdwa.pdf}
\caption{ACCN of CCS services in ablation study for user1 with the continuous stages listed in Table.\ref{tab:user_sequences}. }
\label{fig:ablation_ekdwa}
\end{figure}
\textbf{(1) Combination of Exemplar, Knowledge Distillation, and Weight Aligning}
We conduct the ablation experiment based on user1 at five continuous stages listed in Table.~\ref{tab:user_sequences}. Table.~\ref{tab:ablation_ekdwa} shows that exemplar has a significant impact on reducing model's performance degradation. Without the assistance of exemplar, the accuracy will be 19.29\%,40.9\% and 33.9\% lower at final stage in RFID, Wi-Fi and mmWave modalities respectively, even if we adopt knowledge distillation and weight aligning techniques. Based on exemplar, we apply knowledge distillation and weight aligning techniques separately. As shown in Table.~\ref{tab:ablation_ekdwa}, weight aligning can increase accuracy of model to a greater extend in RFID and Wi-Fi modalities, and to a lower extend in mmWave modality than knowledge distillation with the assistance of exemplar. Moreover, using knowledge distillation, weight aligning and exemplar at the same time can reap better results than using them separately. Fig.~\ref{fig:ablation_ekdwa} illustrates the ACCN curve for ablation study. When we train model using knowledge distillation and weight aligning without exemplar, the ACCN values of models maintain at low values at all stages. In RFID and Wi-Fi modalities, the ACCN values are even lower than values at initial stage. Our method, which utilizes knowledge distillation,weight aligning and exemplar at the same time, keeps on increasing model value as we introduce more unseen classes. Eliminating any techniques in our method can result in lower ACCN values.
\textbf{(2) Loss that applied in Knowledge Distillation}
% {
% \color{red}
We compare performance of knowledge distillation based on different loss functions.Table.~\ref{tab:ab_loss} lists accuracy for user1 at five continuous stages listed in Table.~\ref{tab:user_sequences} when implementing knowledge distillation using Kullback Leibler~(KL) Divergence distance loss, L1 loss, and mean square error loss (MSE).
The results shown in Table.~\ref{tab:ab_loss} indicate that knowledge distillation based on mean square error loss provides best results among these three losses. As Fig.~\ref{fig:ab_loss} illustrated, all three losses can effectively increase model value. Mean-square error loss can receive slightly higher model value than the other two losses in RFID, Wi-Fi and mmWave modalites. Therefore, we select mean-square error loss as the distillation loss of our method.
% }
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{Modality} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C16$\sim$C25} & \textbf{+C26$\sim$C35} & \textbf{+C36$\sim$C45} & \textbf{+C46$\sim$C55} \\
\midrule
\multirow{3}{*}{RFID} & KLD & 72.96 & 52.07 & 40.65 & 35.06 & 29.53\\ & L1 & 74.04 & 52.98 & 38.56 & 33.49 & 29.29\\
& MSE & 75.96 & 58.56 & 44.89 & 38.23 & 34.01\\
\midrule
\multirow{3}{*}{Wi-Fi} & KLD & 95.07 & 80.64 & 73.67 & 60.84 & 61.97\\
& L1 & 95.07 & 81.24 & 74.06 & 61.28 & 63.37\\
& MSE & 95.07 & 83.56 & 75.48 & 64.12 & 65.10 \\
\midrule
\multirow{3}{*}{mmWave} & KLD & 91.11 & 82.0 & 74.32 & 64.69 & 63.29\\
& L1 & 91.11 & 84.24 & 77.67 & 66.89 & 66.70\\
& MSE & 91.11 & 84.31 & 77.84 & 69.19 & 68.54\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Comparative analysis of CCS services' accuracy for user1 with the continuous stages listed in Table.\ref{tab:user_sequences} on different distillation losses, including Kullback Leibler~(KL) Divergence distance loss, L1 loss and mean square error loss.}
\label{tab:ab_loss}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figs/ab_loss.pdf}
\caption{Comparative analysis of CCS services' ACCN for user1 with the continuous stages listed in Table.\ref{tab:user_sequences} on different distillation losses, including Kullback Leibler~(KL) Divergence distance loss, L1 loss and mean square error loss.}
\label{fig:ab_loss}
\end{figure}
% including KL distance loss, L1 loss, and mean square error loss (MSE). The results indicate that knowledge distillation based on mean square error loss achieves a model value closer to the ideal curve path.
\textbf{(3) Normalization Method in Weight Aligning}
% {
% \color{red}
We compare accuracy and ACCN of CCS services for user1 at five continuous stages listed in Table.~\ref{tab:user_sequences} when using L1 norm and L2 norm to implement weight aligning technique. As shown in Table~\ref{tab:ab_norm}, L2 norm outperforms L1 norm at all stages. The accuracy will be 7.84\%,0.39\% and 1.39\% higher at final stage in RFID, Wi-Fi and mmWave modalities respectively when using L2 norm instead of L1 norm.
\begin{table}[t]
\setlength{\tabcolsep}{2pt} % 调整列间距
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llccccc}
\toprule
\textbf{Modality} & \textbf{Method} & \textbf{C1$\sim$C15} & \textbf{+C16$\sim$C25} & \textbf{+C26$\sim$C35} & \textbf{+C36$\sim$C45} & \textbf{+C46$\sim$C55} \\
\midrule
\multirow{2}{*}{RFID} & L1 & 74.04 & 52.31 & 38.37 & 29.98 & 26.17 \\
& L2 & 75.96 & 58.56 & 44.89 & 38.23 & 34.01\\
\midrule
\multirow{2}{*}{Wi-Fi} & L1 & 95.07 & 82.31 & 75.11 & 62.73 & 64.71\\
& L2 & 95.07 & 83.56 & 75.48 & 64.12 & 65.1\\
\midrule
\multirow{2}{*}{mmWave} & L1 & 91.11 & 84.22 & 77.67 & 68.11 & 67.15\\
& L2 & 91.11 & 84.31 & 77.84 & 69.19 & 68.54\\
\bottomrule
\end{tabular}
}
\vspace{2pt}
\caption{Comparative analysis of CCS services' accuracy for user1 with the continuous stages listed in Table.\ref{tab:user_sequences} when using L1 norm and L2 norm to implement weight aligning technique.}
\label{tab:ab_norm}
\end{table}
% \begin{figure}[t]
% \centering
% \includesvg[width=\linewidth]{figs/ab_norm.svg}
% \caption{Comparative analysis of CCS services' ACCN for user1 with the continuous stages listed in Table.\ref{tab:user_sequences} when using L1 norm and L2 norm to implement weight aligning technique}
% \label{fig:ab_norm}
% \end{figure}
Fig.\ref{fig:ab_loss} presents ACCN values of CCS services when using different norms for weight aligning. The ACCN curves in Wi-Fi and mmWave modalities exhibit remarkable similarity, while ACCN curve for L2 norm approaches the ideal curve more closely than ACCN curve for L1 norm in RFID modality. The results indicate that weight aligning technique based on L2 norm can assist in maintaining models' performance better than technique based on L1 norm. Therefore, our method apply L2 norm to implement weight aligning technique.
% \section{Discussion and Future Work}\label{sec:discussion}
% Thus far in the paper, we have introduced CCS, a novel schema tailored to meet users' growing customized needs for wireless sensing, and reported favorable results with extensive experiments. Given that this is a newly introduced topic within the community, we dedicate this section to a thorough discussion of the schema from five perspectives. Our objective is to promote a clearer comprehension and possible refinement of the scheme among our peers.
% \textbf{(1) Acceptance for Accuracy Decline in Old Task}
% This issue concerns the tolerance of users towards the trade-off between meeting new requirements and the decrease in accuracy of old categories. With the increase in the number of new categories, the model's retention of knowledge regarding old categories is inevitably compromised. For users, their focus may lean towards the degree of satisfaction with new requirements, with varying attitudes towards the decline in accuracy. Some users may be more inclined to accept a decrease in accuracy as long as the system meets their new demands. However, others may be more sensitive to accuracy declines. For service providers, there exists a need to balance meeting user demands with maintaining system performance, posing challenges in reconciling system flexibility and stability. If decreases in accuracy adversely affect user experience, it may result in user attrition or negative reputation, thereby impacting the service provider's credibility and business development. Therefore, further in-depth exploration and research on this issue are warranted in future studies.
% \textbf{(2) Providers' Willingness to Transfer Exemplars to Users}
% CCS requires service providers to select a subset of their data, i.e. exemplars, and transfer exemplars to users to overcome the catastrophic forgetting challenge in the incremental service stages. However, data is one of the core interests of model service providers. The providers needs to balance the benefits from the incremental services with the potential risks associated with data disclosure. Typically, when the number of users is particularly large, service providers are more inclined to trade exemplars for service benefits.
% In addition, implementing stricter data security measures is crucial. This may involve encrypting sensitive data, implementing access controls, and conducting regular audits to ensure compliance with data protection regulations. Moreover, the selection method for exemplars is a future research direction; an ideal method would aim to minimize the number of exemplars while maintaining high model performance.
% \textbf{(3) Multiple Model Integration}
% % To meet ongoing wireless sensing demands, the most straightforward approach is to train additional models at each stage. However, directly adopting new models can raise a significant issue: in cases with multiple models, users may find it challenging to assess the reliability of the results produced by each model. Additionally, multiple models require extra storage and necessitate separate computations during inference. This not only increases storage costs but also adds to the computational expense of inference, potentially limiting the system's applicability within resource-constrained devices. Therefore, this straightforward approach is not suitable for the services we proposed in this paper at the moment. Future work could explore integrating multiple models to eliminate ambiguities arising from using several models during inference, and further focus on model compression.
% To meet ongoing wireless sensing demands, the most straightforward approach is to train additional models at each stage. However, directly adopting new models can raise a significant issue: in scenarios involving multiple models, users may struggle to assess the reliability of the results produced by each model. Additionally, managing multiple models requires extra storage and necessitates separate computations during inference. This not only increases storage costs but also adds to the computational expense of inference, potentially limiting the system's applicability, especially within resource-constrained devices. Consequently, this straightforward approach is not suitable for the services proposed in this paper at present. Future work could explore methods for integrating multiple models to address ambiguities during inference and further investigate model compression techniques to enhance efficiency.
% \textbf{(4) Focusing on Model Capability Increasing}
% In this paper, we introduce a novel metric for assessing model capability — ACCN, calculated by multiplying the accuracy of each class by the number of classes. Ideally, model capability should exhibit linear growth with the number of classes. However, the catastrophic forgetting problem prevents the model from achieving excellent accuracy on new class tasks while retaining knowledge of old classes. Moreover, previous research methodologies, e.g. OneFi~\cite{onefi}, mostly focus on the accuracy of new classes, which, according to our evaluation criteria, does not necessarily signify an enhancement in model capability. In contrast, our proposed approach based on continuous learning demonstrates a consistent upward trend in model capability with the increasing number of classes. Thus, our method exhibits significant advantages in comprehensively considering both the learning speed of the model and the retention of knowledge about old classes. Future work could aim to increase model capacity by designing continuous learning methods to support incremental wireless sensing services.
% \textbf{(5) Federated Learning is not Suitable}
% CCS is an approach designed to protect user privacy by ensuring that user data is not uploaded to the service provider. Federated learning can also meet this requirement with a simple design. For example, a user can train a model locally and forwards model updates, e.g. gradients~\cite{mcmahan2017communication}, to the service provider for aggregation, thereby generating a global model without the need to transmit raw data to the service provider. However, some research has also indicated that it is possible to reconstruct the original data using only gradients~\cite{zhu2019deep}.
% In addition, federated learning approach entails considerable network overhead due to the necessity of transmitting model updates between the service provider and users. Moreover, when the number of users is very large, such as in the millions, the network costs will be substantial. More importantly, this will significantly increase the server's computational resource requirements, potentially reducing the service provider’s profit margins. CCS is a resource-light solution for service providers to offer incremental service capabilities, ensuring that user data remains on the device and thereby preserving user privacy.
% CCS is a facilitates service providers to conduct incremental tasks at user terminals,
% In federated learning, Data is distributed across various users. Each location independently trains a model and forwards local updates, e.g. gradients~\cite{}, to a central server for aggregation, thereby generating a global model without the need to transmit raw data to the central server. Although federated learning offers significant advantages in enhancing data privacy, it entails considerable communication overhead due to the necessity of transmitting model updates between the central server and users.
% Additionally, the complexity of model aggregation requires ensuring the performance and accuracy of the global model while merging local model updates.
% Given our practical system requirements and experimental conditions, we have chosen an incremental learning approach to ensure the efficacy of model training and deployment in a more direct manner, while also addressing user privacy concerns.
% We refer to the end-user devices or environments where incremental learning services are deployed, such as smart home systems for human activity recognition, as "terminals." Firstly, our incremental learning task deviates from typical distributed data scenarios, as user updates at the terminal occur in a streaming fashion. Federated learning, however, necessitates training a new model for each new task and subsequently aggregating models, inevitably increasing training complexity and time. Secondly, our system facilitates service providers to conduct incremental tasks at user terminals, ensuring that user data remains on the device and thereby preserving user privacy.
\section{Conclusion}\label{sec:conclusion}
Perceiving wireless sensing as being at the tipping point of transitioning from proof-of-concept to large-scale deployment, we take a step forward by envisioning a service schema. This model entails wireless sensing service providers and users, where providers offer sensing models to users who, in turn, may continuously generate new sensing demands through usage. To enable providers to meet these ongoing user needs, we propose CCS, which facilitates the provision of updated models to satisfy emerging demands. Incorporating techniques like continuous learning, exemplar selection, knowledge distillation, and weight aligning, CCS ensures that new models not only meet new demands but also retain capability for old services. Extensive experiments on the large-scale XRF55 dataset, including sequential introduction of new demands, comparative studies, and ablation experiments, demonstrate the superior sustained service capability of CCS, significantly outperforming existing approaches. Furthermore, we delve into CCS by discussing it from multiple views, including user acceptance, provider incentives, model integration, model capacity, and federated learning.
While CCS may not be the sole or perfect model for future wireless sensing services, it serves as a catalyst for academia and industry to contemplate the paradigms of service or cloud services beyond this tipping point.
% {\BibTeX} does not work by magic. It doesn't get the bibliographic
% data from thin air but from .bib files. If you use {\BibTeX} to produce a
% bibliography you must send the .bib files.
% {\LaTeX} can't read your mind. If you assign the same label to a
% subsubsection and a table, you might find that Table I has been cross
% referenced as Table IV-B3.
% {\LaTeX} does not have precognitive abilities. If you put a
% \verb|\label| command before the command that updates the counter it's
% supposed to be using, the label will pick up the last counter to be
% cross referenced instead. In particular, a \verb|\label| command
% should not go before the caption of a figure or a table.
% Do not use \verb|\nonumber| inside the \verb|{array}| environment. It
% will not stop equation numbers inside \verb|{array}| (there won't be
% any anyway) and it might stop a wanted equation number in the
% surrounding equation.
% \subsection{Some Common Mistakes}\label{SCM}
% \begin{itemize}
% \item The word ``data'' is plural, not singular.
% \item The subscript for the permeability of vacuum $\mu_{0}$, and other common scientific constants, is zero with subscript formatting, not a lowercase letter ``o''.
% \item In American English, commas, semicolons, periods, question and exclamation marks are located within quotation marks only when a complete thought or name is cited, such as a title or full quotation. When quotation marks are used, instead of a bold or italic typeface, to highlight a word or phrase, punctuation should appear outside of the quotation marks. A parenthetical phrase or statement at the end of a sentence is punctuated outside of the closing parenthesis (like this). (A parenthetical sentence is punctuated within the parentheses.)
% \item A graph within a graph is an ``inset'', not an ``insert''. The word alternatively is preferred to the word ``alternately'' (unless you really mean something that alternates).
% \item Do not use the word ``essentially'' to mean ``approximately'' or ``effectively''.
% \item In your paper title, if the words ``that uses'' can accurately replace the word ``using'', capitalize the ``u''; if not, keep using lower-cased.
% \item Be aware of the different meanings of the homophones ``affect'' and ``effect'', ``complement'' and ``compliment'', ``discreet'' and ``discrete'', ``principal'' and ``principle''.
% \item Do not confuse ``imply'' and ``infer''.
% \item The prefix ``non'' is not a word; it should be joined to the word it modifies, usually without a hyphen.
% \item There is no period after the ``et'' in the Latin abbreviation ``et al.''.
% \item The abbreviation ``i.e.'' means ``that is'', and the abbreviation ``e.g.'' means ``for example''.
% \end{itemize}
% An excellent style manual for science writers is \cite{b7}.
% \subsection{Authors and Affiliations}
% \textbf{The class file is designed for, but not limited to, six authors.} A
% minimum of one author is required for all conference articles. Author names
% should be listed starting from left to right and then moving down to the
% next line. This is the author sequence that will be used in future citations
% and by indexing services. Names should not be listed in columns nor group by
% affiliation. Please keep your affiliations as succinct as possible (for
% example, do not differentiate among departments of the same organization).
% \subsection{Identify the Headings}
% Headings, or heads, are organizational devices that guide the reader through
% your paper. There are two types: component heads and text heads.
% Component heads identify the different components of your paper and are not
% topically subordinate to each other. Examples include Acknowledgments and
% References and, for these, the correct style to use is ``Heading 5''. Use
% ``figure caption'' for your Figure captions, and ``table head'' for your
% table title. Run-in heads, such as ``Abstract'', will require you to apply a
% style (in this case, italic) in addition to the style provided by the drop
% down menu to differentiate the head from the text.
% Text heads organize the topics on a relational, hierarchical basis. For
% example, the paper title is the primary text head because all subsequent
% material relates and elaborates on this one topic. If there are two or more
% sub-topics, the next level head (uppercase Roman numerals) should be used
% and, conversely, if there are not at least two sub-topics, then no subheads
% should be introduced.
% \subsection{Figures and Tables}
% \paragraph{Positioning Figures and Tables} Place figures and tables at the top and
% bottom of columns. Avoid placing them in the middle of columns. Large
% figures and tables may span across both columns. Figure captions should be
% below the figures; table heads should appear above the tables. Insert
% figures and tables after they are cited in the text. Use the abbreviation
% ``Fig.~\ref{fig}'', even at the beginning of a sentence.
% \begin{table}[htbp]
% \caption{Table Type Styles}
% \begin{center}
% \begin{tabular}{|c|c|c|c|}
% \hline
% \textbf{Table}&\multicolumn{3}{|c|}{\textbf{Table Column Head}} \\
% \cline{2-4}
% \textbf{Head} & \textbf{\textit{Table column subhead}}& \textbf{\textit{Subhead}}& \textbf{\textit{Subhead}} \\
% \hline
% copy& More table copy$^{\mathrm{a}}$& & \\
% \hline
% \multicolumn{4}{l}{$^{\mathrm{a}}$Sample of a Table footnote.}
% \end{tabular}
% \label{tab1}
% \end{center}
% \end{table}
% \begin{figure}[htbp]
% \centerline{\includegraphics{fig1.png}}
% \caption{Example of a figure caption.}
% \label{fig}
% \end{figure}
% Figure Labels: Use 8 point Times New Roman for Figure labels. Use words
% rather than symbols or abbreviations when writing Figure axis labels to
% avoid confusing the reader. As an example, write the quantity
% ``Magnetization'', or ``Magnetization, M'', not just ``M''. If including
% units in the label, present them within parentheses. Do not label axes only
% with units. In the example, write ``Magnetization (A/m)'' or ``Magnetization
% \{A[m(1)]\}'', not just ``A/m''. Do not label axes with a ratio of
% quantities and units. For example, write ``Temperature (K)'', not
% ``Temperature/K''.
% \section*{Acknowledgment}
% The preferred spelling of the word ``acknowledgment'' in America is without
% an ``e'' after the ``g''. Avoid the stilted expression ``one of us (R. B.
% G.) thanks $\ldots$''. Instead, try ``R. B. G. thanks$\ldots$''. Put sponsor
% acknowledgments in the unnumbered footnote on the first page.
% \section*{References}
% Please number citations consecutively within brackets \cite{b1}. The
% sentence punctuation follows the bracket \cite{b2}. Refer simply to the reference
% number, as in \cite{b3}---do not use ``Ref. \cite{b3}'' or ``reference \cite{b3}'' except at
% the beginning of a sentence: ``Reference \cite{b3} was the first $\ldots$''
% Number footnotes separately in superscripts. Place the actual footnote at
% the bottom of the column in which it was cited. Do not put footnotes in the
% abstract or reference list. Use letters for table footnotes.
% Unless there are six authors or more give all authors' names; do not use
% ``et al.''. Papers that have not been published, even if they have been
% submitted for publication, should be cited as ``unpublished'' \cite{b4}. Papers
% that have been accepted for publication should be cited as ``in press'' \cite{b5}.
% Capitalize only the first word in a paper title, except for proper nouns and
% element symbols.
% For papers published in translation journals, please give the English
% citation first, followed by the original foreign-language citation \cite{b6}.
\bibliographystyle{IEEEtran}
\bibliography{reference.bib}
% \begin{thebibliography}{00}
% \bibitem{b1} G. Eason, B. Noble, and I. N. Sneddon, ``On certain integrals of Lipschitz-Hankel type involving products of Bessel functions,'' Phil. Trans. Roy. Soc. London, vol. A247, pp. 529--551, April 1955.
% \bibitem{b2} J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp.68--73.
% \bibitem{b3} I. S. Jacobs and C. P. Bean, ``Fine particles, thin films and exchange anisotropy,'' in Magnetism, vol. III, G. T. Rado and H. Suhl, Eds. New York: Academic, 1963, pp. 271--350.
% \bibitem{b4} K. Elissa, ``Title of paper if known,'' unpublished.
% \bibitem{b5} R. Nicole, ``Title of paper with only first word capitalized,'' J. Name Stand. Abbrev., in press.
% \bibitem{b6} Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, ``Electron spectroscopy studies on magneto-optical media and plastic substrate interface,'' IEEE Transl. J. Magn. Japan, vol. 2, pp. 740--741, August 1987 [Digests 9th Annual Conf. Magnetics Japan, p. 301, 1982].
% \bibitem{b7} M. Young, The Technical Writer's Handbook. Mill Valley, CA: University Science, 1989.
% \end{thebibliography}
% \vspace{12pt}
% \color{red}
\end{document}
% Generated by IEEEtran.bst, version: 1.14 (2015/08/26)
\begin{thebibliography}{10}
\providecommand{\url}[1]{#1}
\csname url@samestyle\endcsname
\providecommand{\newblock}{\relax}
\providecommand{\bibinfo}[2]{#2}
\providecommand{\BIBentrySTDinterwordspacing}{\spaceskip=0pt\relax}
\providecommand{\BIBentryALTinterwordstretchfactor}{4}
\providecommand{\BIBentryALTinterwordspacing}{\spaceskip=\fontdimen2\font plus
\BIBentryALTinterwordstretchfactor\fontdimen3\font minus \fontdimen4\font\relax}
\providecommand{\BIBforeignlanguage}[2]{{%
\expandafter\ifx\csname l@#1\endcsname\relax
\typeout{** WARNING: IEEEtran.bst: No hyphenation pattern has been}%
\typeout{** loaded for the language `#1'. Using the pattern for}%
\typeout{** the default language instead.}%
\else
\language=\csname l@#1\endcsname
\fi
#2}}
\providecommand{\BIBdecl}{\relax}
\BIBdecl
\bibitem{onefi}
R.~Xiao, J.~Liu, J.~Han, and K.~Ren, ``Onefi: One-shot recognition for unseen gesture via cots wifi,'' in \emph{Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems}, 2021, pp. 206--219.
\bibitem{802}
``Ieee draft standard for information technology -- telecommunications and information exchange between systems local and metropolitan area networks -- specific requirements - part 11: Wireless lan medium access control (mac) and physical layer (phy) specifications-amendment 4: Enhancements for wireless local area network (wlan) sensing,'' \emph{IEEE P802.11bf/D4.0, April 2024}, pp. 1--231, 2024.
\bibitem{soli}
J.~Lien, N.~Gillian, M.~E. Karagozler, P.~Amihood, C.~Schwesig, E.~Olson, H.~Raja, and I.~Poupyrev, ``Soli: Ubiquitous gesture sensing with millimeter wave radar,'' \emph{ACM Transactions on Graphics (TOG)}, vol.~35, no.~4, pp. 1--19, 2016.
\bibitem{fl1}
Z.~Xiao, X.~Xu, H.~Xing, F.~Song, X.~Wang, and B.~Zhao, ``A federated learning system with enhanced feature extraction for human activity recognition,'' \emph{Knowledge-Based Systems}, vol. 229, p. 107338, 2021.
\bibitem{fl2}
K.~Sozinov, V.~Vlassov, and S.~Girdzijauskas, ``Human activity recognition using federated learning,'' in \emph{2018 IEEE Intl Conf on Parallel \& Distributed Processing with Applications, Ubiquitous Computing \& Communications, Big Data \& Cloud Computing, Social Computing \& Networking, Sustainable Computing \& Communications (ISPA/IUCC/BDCloud/SocialCom/SustainCom)}.\hskip 1em plus 0.5em minus 0.4em\relax IEEE, 2018, pp. 1103--1111.
\bibitem{fl3}
X.~Ouyang, Z.~Xie, J.~Zhou, J.~Huang, and G.~Xing, ``Clusterfl: a similarity-aware federated learning system for human activity recognition,'' in \emph{Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services}, 2021, pp. 54--66.
\bibitem{fl4}
B.~Hitaj, G.~Ateniese, and F.~Perez-Cruz, ``Deep models under the gan: information leakage from collaborative deep learning,'' in \emph{Proceedings of the 2017 ACM SIGSAC conference on computer and communications security}, 2017, pp. 603--618.
\bibitem{fl5}
R.~Shokri, M.~Stronati, C.~Song, and V.~Shmatikov, ``Membership inference attacks against machine learning models,'' in \emph{2017 IEEE symposium on security and privacy (SP)}.\hskip 1em plus 0.5em minus 0.4em\relax IEEE, 2017, pp. 3--18.
\bibitem{cf1}
A.~Robins, ``Catastrophic forgetting, rehearsal and pseudorehearsal,'' \emph{Connection Science}, vol.~7, no.~2, pp. 123--146, 1995.
\bibitem{cf2}
R.~M. French, ``Catastrophic forgetting in connectionist networks,'' \emph{Trends in cognitive sciences}, vol.~3, no.~4, pp. 128--135, 1999.
\bibitem{cf3}
J.~Kirkpatrick, R.~Pascanu, N.~Rabinowitz, J.~Veness, G.~Desjardins, A.~A. Rusu, K.~Milan, J.~Quan, T.~Ramalho, A.~Grabska-Barwinska \emph{et~al.}, ``Overcoming catastrophic forgetting in neural networks,'' \emph{Proceedings of the national academy of sciences}, vol. 114, no.~13, pp. 3521--3526, 2017.
\bibitem{ex1}
C.~He, R.~Wang, S.~Shan, and X.~Chen, ``Exemplar-supported generative reproduction for class incremental learning.'' in \emph{BMVC}, vol.~1, 2018, p.~2.
\bibitem{ex2}
Z.~Luo, Y.~Liu, B.~Schiele, and Q.~Sun, ``Class-incremental exemplar compression for class-incremental learning,'' in \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 2023, pp. 11\,371--11\,380.
\bibitem{Herding}
\BIBentryALTinterwordspacing
M.~Welling, ``\BIBforeignlanguage{en-US}{Herding dynamical weights to learn},'' in \emph{\BIBforeignlanguage{en-US}{Proceedings of the 26th Annual International Conference on Machine Learning}}, Jun 2009. [Online]. Available: \url{http://dx.doi.org/10.1145/1553374.1553517}
\BIBentrySTDinterwordspacing
\bibitem{wa}
B.~Zhao, X.~Xiao, G.~Gan, B.~Zhang, and S.-T. Xia, ``Maintaining discrimination and fairness in class incremental learning,'' in \emph{2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, Jun 2020.
\bibitem{xrf55}
F.~Wang, Y.~Lv, M.~Zhu, H.~Ding, and J.~Han, ``Xrf55: A radio frequency dataset for human indoor action analysis,'' \emph{Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies}, vol.~8, no.~1, pp. 1--34, 2024.
\bibitem{icarl}
S.-A. Rebuffi, A.~Kolesnikov, G.~Sperl, and C.~H. Lampert, ``icarl: Incremental classifier and representation learning,'' in \emph{Proceedings of the IEEE conference on Computer Vision and Pattern Recognition}, 2017, pp. 2001--2010.
\bibitem{ucir}
\BIBentryALTinterwordspacing
S.~Hou, X.~Pan, C.~C. Loy, Z.~Wang, and D.~Lin, ``\BIBforeignlanguage{en-US}{Learning a unified classifier incrementally via rebalancing},'' in \emph{\BIBforeignlanguage{en-US}{2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}}, Jun 2019. [Online]. Available: \url{http://dx.doi.org/10.1109/cvpr.2019.00092}
\BIBentrySTDinterwordspacing
\bibitem{bic}
\BIBentryALTinterwordspacing
Y.~Wu, Y.~Chen, L.~Wang, Y.~Ye, Z.~Liu, Y.~Guo, and Y.~Fu, ``\BIBforeignlanguage{en-US}{Large scale incremental learning},'' in \emph{\BIBforeignlanguage{en-US}{2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}}, Jun 2019. [Online]. Available: \url{http://dx.doi.org/10.1109/cvpr.2019.00046}
\BIBentrySTDinterwordspacing
\bibitem{wifringe}
M.~T. Islam and S.~Nirjon, ``Wi-fringe: Leveraging text semantics in wifi csi-based device-free named gesture recognition,'' in \emph{2020 16th International Conference on Distributed Computing in Sensor Systems (DCOSS)}.\hskip 1em plus 0.5em minus 0.4em\relax IEEE, 2020, pp. 35--42.
\bibitem{lwf}
\BIBentryALTinterwordspacing
Z.~Li and D.~Hoiem, \emph{\BIBforeignlanguage{en-US}{Learning without Forgetting}}, Jan 2016, p. 614–629. [Online]. Available: \url{http://dx.doi.org/10.1007/978-3-319-46493-0_37}
\BIBentrySTDinterwordspacing
\bibitem{regular1}
\BIBentryALTinterwordspacing
A.~Chaudhry, P.~K. Dokania, T.~Ajanthan, and P.~H.~S. Torr, \emph{\BIBforeignlanguage{en-US}{Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence}}, Jan 2018, p. 556–572. [Online]. Available: \url{http://dx.doi.org/10.1007/978-3-030-01252-6_33}
\BIBentrySTDinterwordspacing
\bibitem{regular2}
\BIBentryALTinterwordspacing
R.~Aljundi, F.~Babiloni, M.~Elhoseiny, M.~Rohrbach, and T.~Tuytelaars, \emph{\BIBforeignlanguage{en-US}{Memory Aware Synapses: Learning what (not) to forget}}, Jan 2018, p. 144–161. [Online]. Available: \url{http://dx.doi.org/10.1007/978-3-030-01219-9_9}
\BIBentrySTDinterwordspacing
\bibitem{regular3}
S.-W. Lee, J.-H. Kim, J.~Jun, J.-W. Ha, and B.-T. Zhang, ``\BIBforeignlanguage{en-US}{Overcoming catastrophic forgetting by incremental moment matching},'' \emph{\BIBforeignlanguage{en-US}{Cornell University - arXiv,Cornell University - arXiv}}, Mar 2017.
\bibitem{regular4}
D.~Lopez-Paz and M.~Ranzato, ``\BIBforeignlanguage{en-US}{Gradient episodic memory for continual learning},'' \emph{\BIBforeignlanguage{en-US}{Neural Information Processing Systems,Neural Information Processing Systems}}, Jun 2017.
\bibitem{replay1}
\BIBentryALTinterwordspacing
J.~Bang, H.~Kim, Y.~Yoo, J.-W. Ha, and J.~Choi, ``\BIBforeignlanguage{en-US}{Rainbow memory: Continual learning with a memory of diverse samples},'' in \emph{\BIBforeignlanguage{en-US}{2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}}, Jun 2021. [Online]. Available: \url{http://dx.doi.org/10.1109/cvpr46437.2021.00812}
\BIBentrySTDinterwordspacing
\bibitem{replay3}
\BIBentryALTinterwordspacing
D.~Isele and A.~Cosgun, ``\BIBforeignlanguage{en-US}{Selective experience replay for lifelong learning},'' \emph{\BIBforeignlanguage{en-US}{Proceedings of the AAAI Conference on Artificial Intelligence}}, Jun 2022. [Online]. Available: \url{http://dx.doi.org/10.1609/aaai.v32i1.11595}
\BIBentrySTDinterwordspacing
\bibitem{kd1}
G.~Hinton, O.~Vinyals, and J.~Dean, ``\BIBforeignlanguage{en-US}{Distilling the knowledge in a neural network},'' \emph{\BIBforeignlanguage{en-US}{arXiv: Machine Learning,arXiv: Machine Learning}}, Mar 2015.
\bibitem{kd2}
Z.-H. Zhou, Y.~Jiang, and S.~Chen, ``\BIBforeignlanguage{en-US}{Extracting symbolic rules from trained neural network ensembles},'' \emph{\BIBforeignlanguage{en-US}{Ai Communications,Ai Communications}}, May 2003.
\bibitem{rainbow}
\BIBentryALTinterwordspacing
J.~Bang, H.~Kim, Y.~Yoo, J.-W. Ha, and J.~Choi, ``\BIBforeignlanguage{en-US}{Rainbow memory: Continual learning with a memory of diverse samples},'' in \emph{\BIBforeignlanguage{en-US}{2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}}, Jun 2021. [Online]. Available: \url{http://dx.doi.org/10.1109/cvpr46437.2021.00812}
\BIBentrySTDinterwordspacing
\bibitem{gradient}
R.~Aljundi, M.~Lin, B.~Goujaud, and Y.~Bengio, ``\BIBforeignlanguage{en-US}{Gradient based sample selection for online continual learning},'' \emph{\BIBforeignlanguage{en-US}{Neural Information Processing Systems,Neural Information Processing Systems}}, Jan 2019.
\bibitem{anchor}
\BIBentryALTinterwordspacing
A.~Chaudhry, A.~Gordo, P.~Dokania, P.~Torr, and D.~Lopez-Paz, ``\BIBforeignlanguage{en-US}{Using hindsight to anchor past knowledge in continual learning},'' \emph{\BIBforeignlanguage{en-US}{Proceedings of the AAAI Conference on Artificial Intelligence}}, p. 6993–7001, Sep 2022. [Online]. Available: \url{http://dx.doi.org/10.1609/aaai.v35i8.16861}
\BIBentrySTDinterwordspacing
\bibitem{he2016deep}
K.~He, X.~Zhang, S.~Ren, and J.~Sun, ``Deep residual learning for image recognition,'' in \emph{Proceedings of the IEEE conference on computer vision and pattern recognition}, 2016, pp. 770--778.
\bibitem{wang2019temporal}
F.~Wang, Y.~Song, J.~Zhang, J.~Han, and D.~Huang, ``Temporal unet: Sample level human action recognition using wifi,'' \emph{arXiv preprint arXiv:1904.11953}, 2019.
\end{thebibliography}