PaperlayerView exact revision on arXiv

Source-linked AI summary

Transfer Learning for Future Wireless Networks: A Comprehensive Survey

Cong T. Nguyen, Nguyen Van Huynh, Nam H. Chu, Yuris Mulya Saputra, Dinh Thai Hoang, Diep N. Nguyen, Quoc-Viet Pham, Dusit Niyato, Eryk Dutkiewicz, Won-Joo Hwang

arXiv:2102.07572v2cs.LGcs.NI

TL;DR

Conventional wireless ML faces scarce data, shifting environments, long training, and limited device capacity. This paper surveys TL foundations and wireless applications; in spectrum sensing, reported TL gains reach up to 94% in training time and 20% in sensing time.

  • Problem

    Conventional wireless ML is challenged by scarce training data, changing environments, long training times, and limited device computing capacity.

    Source: §I · p. 1

  • Method

    The paper provides a tutorial and survey of TL techniques and their applications to wireless-network problems, then discusses open issues and future directions.

    Source: §I · p. 2, §X · p. 35

  • Results

    In spectrum sensing, TL improved CNN training time by up to 94% and sensing time by 20%.

    Source: 1) Spectrum Sensing · p. 12

  • Takeaways & Limitations

    The survey covers TL applications in spectrum management, localization, signal recognition, security, human activity recognition, and caching.

    Source: §X · p. 35

Abstract

from arXiv · show

With outstanding features, Machine Learning (ML) has been the backbone of numerous applications in wireless networks. However, the conventional ML approaches have been facing many challenges in practical implementation, such as the lack of labeled data, the constantly changing wireless environments, the long training process, and the limited capacity of wireless devices. These challenges, if not addressed, will impede the effectiveness and applicability of ML in future wireless networks. To address these problems, Transfer Learning (TL) has recently emerged to be a very promising solution. The core idea of TL is to leverage and synthesize distilled knowledge from similar tasks as well as from valuable experiences accumulated from the past to facilitate the learning of new problems. Doing so, TL techniques can reduce the dependence on labeled data, improve the learning speed, and enhance the ML methods' robustness to different wireless environments. This article aims to provide a comprehensive survey on applications of TL in wireless networks. Particularly, we first provide an overview of TL including formal definitions, classification, and various types of TL techniques. We then discuss diverse TL approaches proposed to address emerging issues in wireless networks. The issues include spectrum management, localization, signal recognition, security, human activity recognition and caching, which are all important to next-generation networks such as 5G and beyond. Finally, we highlight important challenges, open issues, and future research directions of TL in future wireless networks.

I. INTRODUCTION

Wireless ML is challenged by changing environments, scarce data, long training, device constraints, and centralized-data privacy concerns. The paper introduces TL and surveys its foundations and wireless applications.

I. INTRODUCTION

  • Changing wireless conditions, scarce training data, long training times, limited device capacity, and privacy concerns constrain conventional ML in wireless networks.Centralized training can also require collecting raw data at servers.Source: §I · p. 1
  • TL transfers knowledge from similar tasks or past experience to improve learning on a new problem.The paper summarizes benefits including richer training data, faster learning, reduced target-side computing and communication demands, and privacy protection.Source: §I · pp. 1–2, A. Overview of Transfer Learning · p. 3
  • The paper surveys TL foundations, application approaches for wireless-network challenges, and open issues and future directions.Its application coverage includes spectrum management, localization, signal recognition, security, human activity recognition, and caching.Source: §I · pp. 2–3

B. Fundamentals of Transfer Learning

TL transfers knowledge from source domains and tasks to improve a target predictive function when source and target differ. Its use requires choices about what, when, and how to transfer.

B. Fundamentals of Transfer Learning

  • A domain comprises a feature space and its marginal distribution, while a task comprises a label space and a predictive function.The target domain often has far fewer instances than the source domain.Source: B. Fundamentals of Transfer Learning · p. 3
  • TL learns the target predictive function using source-domain and source-task knowledge when the source and target domains or tasks differ.Differences may involve feature spaces, marginal distributions, label spaces, or conditional label distributions.Source: B. Fundamentals of Transfer Learning · pp. 3–4

C. Transfer Learning Techniques

  • Transfer design asks what knowledge to transfer, when transfer is appropriate, and how to implement it.Unrelated source and target domains can make transfer ineffective or cause negative transfer.Source: C. Transfer Learning Techniques · p. 4
  • The survey categorizes TL by transferred knowledge as feature-, parameter-, relational-, or instance-based, and by setting as inductive, transductive, or unsupervised.The settings distinguish source and target task/domain relationships and the availability of labeled data.Source: C. Transfer Learning Techniques · p. 4

1) Inductive Transfer Learning:

The survey distinguishes inductive, transductive, and unsupervised TL by source-target task and domain relationships and label availability, and describes deep TL as transferring learned model knowledge.

1) Inductive Transfer Learning:

  • Inductive TL transfers knowledge between different tasks and requires labeled target data for good target-function performance.It includes self-taught learning, which uses unlabeled source data to form feature representations, and multi-task learning, where both domains have labels and learn simultaneously.Source: 1) Inductive Transfer Learning · p. 4

2) Transductive Transfer Learning:

  • Transductive TL addresses different source and target domains with the same task, requiring unlabeled target data during training.The domains may differ in feature spaces or, with the same feature space, in marginal distributions.Source: 2) Transductive Transfer Learning · p. 5

3) Unsupervised Transfer Learning:

  • Unsupervised TL transfers between different tasks when labeled data are unavailable in both domains.The survey describes source-target clustering to find common features and a self-supervised approach that learns a discriminative target-data space.Source: 3) Unsupervised Transfer Learning · p. 5

D. Deep Transfer Learning

  • Deep TL treats a trained model’s architecture and parameters as knowledge that can be transferred across domains.The survey identifies off-the-shelf pre-trained models, pre-trained models as feature extractors, and fine-tuning as three typical strategies.Source: D. Deep Transfer Learning · pp. 5–6

1) Deep Transfer Learning Strategies:

The survey describes three ways to reuse pre-trained deep models: direct use, feature extraction, and fine-tuning. Their suitability depends on source-target task and domain similarity and the amount of target data.

1) Deep Transfer Learning Strategies:

  • Off-the-shelf pre-trained models can be used directly, but require the source and target domains to match and tasks to be similar.When tasks differ substantially, direct use may not achieve good results; the survey illustrates this with car-type classification versus car detection.Source: 1) Deep Transfer Learning Strategies · p. 6
  • A pre-trained model can extract target-data features that encode source-domain knowledge for a downstream learner.Lower layers capture more general features, while deeper layers capture more task- and domain-specific features; feature depth should reflect source-target task similarity.Source: 1) Deep Transfer Learning Strategies · pp. 6–7
  • Fine-tuning continues training some or all pre-trained parameters with target data, using weight initialization or selective layer training.With small target datasets and large models, freezing more layers can help prevent overfitting; with larger target data and fewer parameters, more layers can be trained.Source: 1) Deep Transfer Learning Strategies · p. 7

2) How to select the best DTL strategy:

DTL strategy selection depends especially on target-data size and source–target similarity. The section also reviews how DRL handles large state spaces and stabilizes learning, then contrasts value-based and policy-based methods.

2) How to select the best DTL strategy:

  • For small labeled targets, use a pretrained feature extractor when source data are similar, transfer general early-layer features when they differ, and prefer weight initialization when target data are large.The survey cautions that fine-tuning on a small similar target set may overfit.Source: 2) How to select the best DTL strategy · p. 8

E. Transfer Learning in Deep Reinforcement Learning

1) Deep Reinforcement Learning:

  • DRL uses DNNs to represent value functions or policies, addressing the difficulty of tabular Q-learning in environments with many states and actions.Value-based methods estimate a Q-function; policy-based methods learn a policy directly, which avoids calculating Q-values for very large or continuous action spaces.Source: 1) Deep Reinforcement Learning · pp. 8–9
  • Experience replay trains on randomly sampled past transitions to loosen data correlation, while a periodically copied target Q-network makes learning more stable.The Q-network is trained at each time step; target-network parameters are copied at predefined intervals.Source: 1) Deep Reinforcement Learning · p. 9

2) Transfer Learning Strategies in Deep Reinforcement Learning:

TL-DRL transfers source-MDP knowledge to train a policy for a target MDP. The section organizes approaches by the knowledge transferred, including rewards, demonstrations, policies, mappings, and representations.

Transfer Learning Strategies in Deep Reinforcement Learning:

  • TL-DRL trains a target-MDP policy using source knowledge—data, policy, and environment dynamics—together with target information.The target policy is estimated by a DNN trained on both source and target data.Source: 2) Transfer Learning Strategies in… · pp. 9–10
  • Reward shaping adds a source-informed reward to the target reward to guide action selection, with minimal changes to the DRL framework.The modified reward is RRS = R + F.Source: 2) Transfer Learning Strategies in… · p. 10
  • Demonstration-based transfer uses source-MDP transition examples to guide exploration, either by initializing target components offline or directing actions during online learning.Demonstrations may come from optimal, near-optimal, or suboptimal policies.Source: 2) Transfer Learning Strategies in… · p. 10
  • Policy transfer distills source policies into a target policy by matching action distributions or directly reuses source policies according to their target-MDP performance.Policy reuse also includes the current target policy among the policies that may be selected.Source: 2) Transfer Learning Strategies in… · p. 10
  • Inter-task mapping learns correspondences between source and target MDP states or transition representations, but transition mapping often assumes similarity in dynamics and state representations.State mappings can reshape the target reward; the described approaches use autoencoders for states and sparse pseudo-input Gaussian processes for transitions.Source: 2) Transfer Learning Strategies in… · p. 10
  • Representation reuse transfers state or action features directly or through task-invariant features; progressive networks grow with tasks, while modular designs keep a fixed size.Modular designs can separate task-specific and environment-specific components.Source: 2) Transfer Learning Strategies in… · pp. 10–11

3) How to select the best TL-DRL strategy:

Selecting a TL-DRL scheme requires identifying how the source and target MDPs differ and checking whether a method restricts the RL algorithm it can use.

3) How to select the best TL-DRL strategy:

  • Choose a TL-DRL scheme by assessing differences in state space, action space, reward, and environment dynamics, and checking algorithm compatibility.The cited Deep Q-learning from Demonstration framework applies only to Deep Q-Learning.Source: 3) How to select the best TL-DRL… · p. 11

III. APPLICATIONS OF TRANSFER LEARNING FOR SPECTRUM MANAGEMENT

Spectrum-management TL addresses radio variation and short training windows by reusing knowledge across related scenarios. In cognitive radio networks, the surveyed approaches apply this idea to spectrum sensing, channel selection, and interference management.

III. APPLICATIONS OF TRANSFER LEARNING FOR SPECTRUM MANAGEMENT

  • Spectrum-management TL targets changing radio environments and time-sensitive tasks where conventional ML models trained for one scenario may be less effective.The section identifies resource allocation, channel estimation, and spectrum sensing as relevant tasks.Source: §III · p. 11

A. Cognitive Radio Networks

  • In cognitive radio networks, TL draws on related scenarios to improve the effectiveness and robustness of ML for sensing, handoff prediction, and interference management.These tasks otherwise require substantial training time and data in changing radio environments.Source: A. Cognitive Radio Networks · p. 11

1) Spectrum Sensing:

  • TL-based spectrum sensing transfers pretrained CNNs across signal types or environments, improving detection metrics in the reported studies.The approach in fine-tunes on target data and improves detection probability and false-alarm rate by up to 25%; reports up to 30% higher detection probability at low SNR. With limited target data, can perform worse than energy detectors.Source: 1) Spectrum Sensing · p. 11, Figure 8 · p. 12

2) Channel Selection:

  • For channel selection, TL reuses spectrum-band data or other users’ learned policies to improve prediction or speed new users’ learning.Using data from other bands, reports up to 2.5-times higher prediction accuracy, especially with limited target data; transfer is effective only for certain band pairs, requiring similarity analysis. Approaches in and improve convergence by up to 30% and 14%, respectively.Source: 2) Channel Selection · p. 12

3) Interference Management:

  • For interference management, transferring Q-tables or nearby users’ learned models improves learning or reduces the risk of low macrocell capacity.[67] reports up to a 20-times reduction in the probability that macrocell capacity falls below a threshold; and report learning-time reductions of approximately 25% and 80%, respectively.Source: 3) Interference Management · p. 13

4) Radio Map Construction:

The section surveys TL applications in radio-map construction and resource allocation. Across these examples, approaches transfer learned models, features, or policies to address data scarcity, changing environments, or long training times.

4) Radio Map Construction:

  • For antenna-tilt radio maps, transferring the first two feed-forward NN layers gives the best reported strategy, with prediction accuracy up to 12% higher than KNN and random forest.A separate GAN-based approach transfers radio-environment features from other regions and reconstructs power-spectrum maps more accurately than traditional methods without TL, though the TL effect was not comprehensively evaluated.Source: 4) Radio Map Construction · p. 13, Figure 9 · p. 13

B. Resource Allocation

  • Resource allocation spans decisions such as spectrum, resource-block, and power allocation, which can improve network operation efficiency and user experience.The section notes that traditional ML models may perform poorly with scarce data and require long training.Source: B. Resource Allocation · pp. 13–14

1) Channel Assignment:

  • For channel assignment, transferring Q-values across user clusters reduces delay and transmission failure rate by up to 1.5 times versus the approach without TL.The approach clusters users by RSRQ and uses transmission success or failure to update cluster Q-values.Source: 1) Channel Assignment · p. 14, Figure 10 · p. 14

2) Power Allocation and Energy Efficiency:

  • For power allocation and energy efficiency, surveyed TL methods transfer previous policies or model weights to improve learning or energy efficiency; outcomes vary by approach and environmental change.Reported energy-efficiency gains were up to 22% and 60%. In, learning sped up by 37.5% under moderate environmental change, but TL was ineffective under major changes.Source: 2) Power Allocation and Energy… · p. 14

3) Resource Block Allocation:

  • For resource-block allocation, TL approaches improve packet reception ratio or accelerate learning in new environments.The impact of TL itself is not evaluated in [80]; reports an 18.2% learning-speed improvement.Source: 3) Resource Block Allocation · p. 15

4) Resource Utilization Prediction:

  • For resource-utilization prediction, transferring a pretrained LSTM and fine-tuning its later layers reduces target-model training time by approximately half.The approach fixes the first layers, which are described as capturing more general features.Source: 4) Resource Utilization Prediction · p. 15

C. Channel Estimation and Prediction

Transfer learning is applied to channel prediction and spectrum-management problems where changing environments, scarce data, and training costs constrain conventional ML. The section also motivates localization TL as a way to reduce repeated signal-data collection and recalibration.

C. Channel Estimation and Prediction

  • For cellular-service traffic prediction, TL improves the reported models’ RMSE and MAE by up to 17.1% and 12.9%, respectively.The approaches transfer a model from one service to another; one uses timestamp data and the other does not.Source: C. Channel Estimation and Prediction · p. 15
  • Transferring more RF-model hyperparameters reduces training time but may worsen prediction accuracy when environments differ.The compared transfers include all hyperparameters, lengthscales and noise hyperparameters, or lengthscales alone.Source: C. Channel Estimation and Prediction · p. 15
  • With no target data, CNN RMSE is 6% lower than LSTM RMSE; with more data, LSTM RMSE is 3% lower than CNN RMSE.The compared CNN and LSTM transfer models predict channel quality indicators across frequency bands or cells.Source: C. Channel Estimation and Prediction · pp. 15–16

D. Other Issues in Spectrum Management

  • UAV path-loss and delay-spread TL achieves lower RMSE than other approaches, while femtocell-model output aggregation improves SVM prediction accuracy by up to 20%.The UAV approaches transfer models across frequencies or scenarios; the femtocell method weights model outputs by data similarity.Source: D. Other Issues in Spectrum Management · p. 16
  • Spectrum-management TL depends on source and target tasks sharing features, yet many reviewed works do not analyze their domains or justify which model components to transfer.Reviewed strategies differ between transferring early layers and transferring the whole source model.Source: D. Other Issues in Spectrum Management · p. 16

IV. APPLICATIONS OF TRANSFER LEARNING FOR LOCALIZATION

  • Localization TL can reduce frequent signal-data re-collection and recalibration in time-varying wireless environments.Indoor localization can use RSS, CSI, or sensor signals to estimate object positions.Source: §IV · p. 16

A. Indoor Localization

  • Indoor localization matches online wireless-signal patterns to an offline signal-to-location map, which can become outdated as indoor signals change.TL approaches are presented as a way to address map staleness without frequent signal-data updates.Source: A. Indoor Localization · p. 17

1) RSS Map-Based Localization:

RSS-map localization transfers or adapts signal-to-location knowledge across time, devices, and environments, while CSI-based approaches use signal fingerprints to address changing conditions. The reviewed methods include map adaptation, reduced-fingerprint schemes, and approaches that avoid repeated data collection.

1) RSS Map-Based Localization:

  • For indoor RSS localization, jointly addressing time and device variation improves accuracy over non-TL methods, while fuzzy clustering with manifold-alignment map updates improves mean location error by up to 71% over.The joint time-and-device approach uses manifold alignment; the clustering approach updates the RSS map to reduce environmental effects.Source: 1) RSS Map-Based Localization · p. 18
  • A map-free inductive scheme transfers RSS knowledge to a few fingerprints and unlabeled observations, reducing mean localization error by up to 20% versus a simulated radio-map scenario.It uses manifold alignment to exploit spatial correlation among observations for direct localization.Source: 1) RSS Map-Based Localization · p. 18
  • Clustering-based KNN/WKNN TL reaches 85% localization accuracy with up to 30% less offline training overhead than non-TL.The method learns source-domain distance metrics and reshapes target-domain distances using the RSS map.Source: 1) RSS Map-Based Localization · p. 18
  • TKL-WinSMS improves localization accuracy by up to 26% over SVR without TKL, while instance weighting improves accuracy by up to 18% or 10% under the two reported comparisons.TKL uses unlabeled target RSS data with labeled source data; instance weighting selects old RSS samples matching the current test distribution.Source: 1) RSS Map-Based Localization · pp. 18–19
  • DMCA aligns mean embeddings and second-order statistics across domains, achieving up to 2.76 meters lower average error distance than non-DL-based learning without fingerprint or hardware calibration.The evaluation used smartphones across different days and targets dynamic environments and heterogeneous hardware.Source: 1) RSS Map-Based Localization · p. 19
  • HDTL maps homogeneous and heterogeneous features from multiple source domains into a common feature space to localize despite target-domain data loss.The review reports higher localization accuracy, measured as lower average error distance, than conventional TL using a stand-alone source domain.Source: 1) RSS Map-Based Localization · p. 19

2) CSI Map-Based Localization:

  • CSI-map localization with CRISLoc maintains a mean error within 0.29 meters and variance up to 8.6 cm when two AP locations change.It re-models high-dimensional CSI fingerprints using an obsolete database and updated measurements.Source: 2) CSI Map-Based Localization · p. 19
  • A device-free passive CSI approach reports mean localization error up to 1.5 meters without re-collecting CSI fingerprints.Deep TL learns hidden fingerprint representations to reduce distribution discrepancy between training and testing samples.Source: 2) CSI Map-Based Localization · pp. 19–20

3) Sensor-Based Localization:

The reviewed localization approaches transfer sensor, RSS, or learned model knowledge across settings, including mobile-robot and large-area outdoor localization. The signal-recognition discussion frames TL as a response to distribution changes and costly data collection and processing.

3) Sensor-Based Localization:

  • iLoom transfers outdoor movement knowledge to indoor dead reckoning and uses an acceleration range box to filter incorrect accelerations, improving localization accuracy by up to 60%.The review attributes the result to indoor and outdoor experiments and reports low-complexity training versus the conventional method without the range box.Source: 3) Sensor-Based Localization · p. 20
  • A CNN-based robot-localization scheme transfers image features to classify positions on a topological map, reaching up to 100% accuracy in the reported apartment setting.The images were captured using GoPro and omnidirectional cameras; the reported 100% result is specifically associated with GoPro images.Source: 3) Sensor-Based Localization · p. 20

B. Outdoor Localization

  • Outdoor CTTL trains location knowledge over a supercell and adapts model weights for smaller cells, reaching 95% positioning accuracy with low communication cost and high robustness.Beacon nodes communicate within their cell to estimate pending-node locations under changing node density and SNR.Source: B. Outdoor Localization · pp. 20–21
  • Outdoor LoRaWAN TL improves location accuracy by up to 75% over SVM after 5000 epochs, while inter-technology RSS-map transfer improves accuracy by up to 16% over traditional ML without knowledge transfer.The approaches use labeled-unlabeled location correlation and transfer between LoRaWAN and wireless M-Bus, respectively.Source: B. Outdoor Localization · p. 21
  • Most reviewed localization work uses inductive TL across similar environments; cross-environment transductive or unsupervised TL and sensor-to-RSS/CSI transfers remain directions for investigation.The review also notes inductive DTL for outdoor positioning with few labeled and many unlabeled data.Source: B. Outdoor Localization · p. 21

V. APPLICATIONS OF TRANSFER LEARNING FOR SIGNAL CLASSIFICATION AND MODULATION RECOGNITION

  • Signal-recognition DL assumes identically distributed data despite receiver sampling-rate and hardware differences, and its training and data-processing demands can be costly or infeasible.The section presents TL as a promising solution for these signal-classification and modulation-recognition challenges.Source: §V · pp. 21–22

A. Signal Classification

Transfer learning reuses models, features, or past experience to address signal-recognition challenges across changing channels, noise, sampling rates, and limited labeled data. The surveyed security applications likewise transfer knowledge across devices, environments, or time to support intrusion detection.

A. Signal Classification

  • Transfer learning addresses channel and hardware variation by reusing source-network layers and fine-tuning them for target settings, with backscatter detection approaching optimal detection with perfect channel-state information.Over-the-air classification using synthetic-data pretraining achieves accuracy close to training on synthetic data; weak-signal backscatter detection uses offline training and online detection.Source: A. Signal Classification · p. 22

B. Modulation Recognition

  • Transferring high-SNR weights can address modulation-recognition accuracy falling below 40% in unstable channels, while other surveyed methods use adversarial transfer or shared features to improve recognition.A semi-supervised CNN and convolutional auto-encoder approach also outperforms conventional approaches and approaches the accuracy of good-channel conditions.Source: B. Modulation Recognition · p. 23

C. Signal Recognition for Radar Systems

  • 98.89% with AlexNet and 99.72% with VGGNet were reported for waveform recognition in noisy environments; radar recognition also showed greater robustness and accuracy than SVM, K-means, and fuzzy-theory methods.A separate radar approach transfers a pretrained model to a different sample rate and outperforms manual-feature and DNN-based methods.Source: C. Signal Recognition for Radar Systems · p. 23

VI. APPLICATIONS OF TRANSFER LEARNING FOR WIRELESS SECURITY

  • For wireless security, transfer learning is applied to intrusion detection, jamming mitigation, anti-eavesdropping encryption, and spam-call filtering to address data, device, environment, and learning-speed challenges.The surveyed security applications include transferring knowledge from past experience and adapting models to device- or user-specific data.Source: §VI · p. 24

A. Intrusion Detection

  • In intrusion detection, source-trained and target-trained autoencoders align layers using MMD, while transferring the first 12 CNN layers improves intersection over union by up to 21%.Transferring environmental-effect knowledge among nearby devices lowers false positives; device-to-device transfer improves detection performance by 70%. Longer-term fingerprint transfer improves recognition but takes more time.Source: A. Intrusion Detection · pp. 24–25

B. Jamming Attacks

The surveyed security applications use transfer learning to reuse prior experience or models when wireless conditions, users, or data availability change. This section covers anti-jamming, other security tasks, and activity-recognition motivation.

B. Jamming Attacks

  • Unsupervised transfer initializes Q-learning from prior experience or another region’s Q table, and the approaches outperform non-transfer learning in reported anti-jamming measures.The WBAN approach reports better BER, SINR, and long-term jammer utility; the multi-region approach converges faster and achieves better BER.Source: B. Jamming Attacks · p. 26

C. Other Security Issues

  • For eavesdropping protection, prior action rewards are incorporated into actor-critic learning; for spam calls, a baseline model is fine-tuned with user-specific call data, reaching up to 99.88% validation accuracy.The survey notes that baseline models require long training on large generalized datasets, while transfer-parameter choices such as transfer rates are often not analyzed.Source: C. Other Security Issues · pp. 26–27

VII. APPLICATIONS OF TRANSFER LEARNING FOR ACTIVITY RECOGNITION

  • Wireless activity-recognition models may be environment-specific, so transfer learning is proposed to support recognition across settings without large computing resources.The applications mentioned include human activity, handwritten signature, and mobile-device recognition.Source: §VII · p. 27

A. Human Activity Recognition

Wireless activity-recognition models can be costly to rebuild for new environments, and CSI or wearable-sensor data provide inputs for transfer-learning approaches. The surveyed methods span CSI-based recognition and sensor-based systems.

1) CSI Signal-Based Human Activity Recognition:

  • Using few labeled CSI samples, WiSign transfers waveform knowledge across users for sign recognition, while CSI-to-image transfer for gesture recognition reports up to 100% accuracy.WiSign outperforms classical SVM in mean prediction accuracy; the gesture system uses a pretrained CNN and fine-tuning.Source: 1) CSI Signal-Based Human Activity… · p. 27
  • TL-Fall transfers falling-velocity knowledge from labeled data in an old condition to a new condition with few labeled samples, outperforming baseline methods in fall-detection sensitivity.It extracts frequency components from CSI amplitudes across Wi-Fi OFDM subcarriers.Source: 1) CSI Signal-Based Human Activity… · p. 27
  • CSI-based TL-HAR improves classification accuracy by up to 3% for single MIMO links and 5.6% for multiple links, while a location-independent approach exceeds 90% accuracy across six activities.The location-independent method can recognize new activities without training samples or with only a few samples.Source: 1) CSI Signal-Based Human Activity… · pp. 27–28

2) Sensor-Based Human Activity Recognition:

  • Wearable-sensor transfer at edge servers improves abnormal-activity prediction accuracy by up to 28%, while collaborative multi-view learning improves activity recognition by up to 9.3% to 83.7%.The latter compares refined source-and-target views with direct reuse of the source model at the target view.Source: 2) Sensor-Based Human Activity… · pp. 28–29
  • Smartphones can train local partial models and aggregate them into a complete model for transfer, achieving accuracy comparable to centralized learning while reducing wireless overhead.This approach uses locally trained sensor data rather than sending raw measurements to the edge data center.Source: 2) Sensor-Based Human Activity… · p. 29
  • FedHealth aggregates encrypted local wearable models and adapts common CNN layers for users, improving accuracy by up to 5.3% on the UCI Smartphone dataset and 21.6% on a Parkinson’s disease dataset.User-specific fully connected layers are updated using a correlation-alignment layer.Source: 2) Sensor-Based Human Activity… · p. 29

3) Probe Response Signal-Based Human Activity Recognition:

The surveyed work applies TL to probe-response activity recognition, Wi-Fi signature recognition, and wireless caching. Across these cases, knowledge from signals, users, or prior domains is reused for recognition or content-popularity prediction.

3) Probe Response Signal-Based Human Activity Recognition:

  • Up to 7% classification-accuracy improvement, with significant training-time reduction, is reported for inductive TL fine-tuning of a Doppler CNN on six probe-response activities.The method uses Wi-Fi probe-response signals without connecting to the actual Wi-Fi network.Source: 3) Probe Response Signal-Based Human… · p. 29

B. Handwritten Signature Recognition

  • The Wi-Fi signature method in transfers a feature extractor across positions using kernel and range space projection learning, reaching 96.75% accuracy with faster training computation.Only the classifier is retrained because signature signals from different positions share a data domain.Source: B. Handwritten Signature Recognition · p. 30

VIII. APPLICATIONS OF TRANSFER LEARNING FOR WIRELESS CACHING

  • Caching TL addresses costly data collection and sparse user-content matrices; individual caching transfers knowledge separately to SBSs without cached-content sharing.Content-popularity prediction supports proactive caching at edge servers near requesting users.Source: §VIII · p. 30, A. Individual Caching · p. 30

1) Content Popularity Prediction at Small BSs:

  • Up to 22% improvement in satisfaction ratio and backhaul offloading gain is reported for inductive TL caching at SBSs versus random caching and collaborative filtering.The evaluation considers different storage sizes, traffic intensities, and backhaul capacities.Source: 1) Content Popularity Prediction at… · p. 30

2) Content Popularity Prediction at Macro BSs:

  • Macro-BS popularity prediction centralizes estimation, and TL transfers user-content patterns; training time falls linearly as source-domain samples increase.The source and target examples include user preferences for different movie genres and users’ request patterns.Source: 2) Content Popularity Prediction at… · pp. 30–31
  • The DQN, imitation-learning, and unsupervised-TL video-rate mechanism outperforms DNNs and random allocation in average user reward, a measure of streaming QoS.TL transfers and fine-tunes DQN parameters to map video rates from source to target domains.Source: 2) Content Popularity Prediction at… · p. 31

B. Cooperative Caching

Cooperative caching applies TL between edge servers to predict requests and place content, while the survey identifies unresolved choices in source tasks, transferred layers, parameters, and fine-tuning. The caching studies also assume popularity remains constant over time.

B. Cooperative Caching

  • Up to 177% cache-hit-rate improvement is reported for cooperative TL caching; RSVD-based CF plus TL reduces request delay and improves user satisfaction.The cooperative framework trains popularity models across edge servers and derives content placement with a greedy algorithm.Source: B. Cooperative Caching · p. 32
  • The surveyed caching works assume a time-constant popularity matrix, although mobile users’ preferences may change and require evaluation of TL robustness.The stated goal is to maintain high cache-hit ratio and low content-request delay over time.Source: B. Cooperative Caching · p. 32

1) Determine the Source Tasks in Wireless Networks:

  • Source choice is difficult because wireless settings offer many candidate domains; weakly related source-target tasks can make TL worse than target-only training.The survey notes that most current works do not adequately justify their source choices.Source: 1) Determine the Source Tasks in… · p. 32

2) Determine What to Transfer in Wireless Networks:

  • Earlier model layers tend to encode more general features and later layers more domain-specific ones; transfer depth depends on source-target similarity and application.The survey notes that few works address how many layers to transfer beyond simulation-based testing.Source: 2) Determine What to Transfer in… · p. 32

3) Determine the TL Parameters in Wireless Networks:

  • Reported best transfer rates conflict: 0.8 beats 0.2 and 0.5 in, while 0.5 is preferred in, underscoring application dependence.Most surveyed approaches transfer all weights without explaining that choice.Source: 3) Determine the TL Parameters in… · pp. 32–33

4) Trade-off between Performance and Training Time:

  • Fine-tuning more layers can improve performance but costs training time; with limited, unrepresentative target labels, fine-tuning all layers may overfit.The appropriate number of layers depends on the application’s training window and data availability.Source: 4) Trade-off between Performance and… · p. 33

1) Beamforming:

The survey reviews transfer-learning approaches for wireless-network challenges, from beamforming and resource optimization to decentralized sharing, and identifies these applications as research directions for future networks.

1) Beamforming:

  • For beamforming, transferred CNNs are fine-tuned for new settings, while freezing all but the final layer can reduce overfitting and improve convergence.A separate CNN-based approach is described without simulations or experiments evaluating its transfer process.Source: 1) Beamforming · p. 33

2) Embedded IoT Devices:

  • Embedded-device learning combines edge/cloud training and parameter reduction with FPGA fine-tuning and periodic model updates, while device-learned information returns to the edge/cloud.The proposed learning and transfer procedure is described as a hybrid of transfer and federated learning.Source: 2) Embedded IoT Devices · pp. 33–34

3) General Optimization Problem in Wireless Networks:

  • Self-imitation transfer speeds branch-and-bound training by using a transferred policy to explore a new optimization task and create labels from branch objective values.The reported training speed-up is 9 to 15 times; unlike fine-tuning, this approach does not require labeled data.Source: 3) General Optimization Problem in… · p. 34

4) Network Selection:

  • Context-aware transfer uses traffic, location, and time information to select among LTE, WLAN, and VLC, achieving higher average reward and faster convergence than conventional reinforcement learning.The approach periodically updates load-statistics distributions as transferred knowledge.Source: 4) Network Selection · p. 34

5) Improving TL with other ML Techniques:

  • For downlink CSI prediction, meta-learning aggregates weights trained separately across datasets and reduces NMSE by up to 15% versus transfer learning and 50% versus no transfer.The alternative transductive transfer approach may overfit when target data are insufficient.Source: 5) Improving TL with other ML Techniques · p. 34

6) Decentralized TL and Blockchain:

  • A blockchain-based decentralized transfer framework supports secure model, data, and resource sharing and automatic payments, with model sharing reported at less than one minute while ensuring security and privacy.The approach addresses decentralized knowledge sharing from sources such as neighboring networks or users.Source: 6) Decentralized TL and Blockchain · pp. 34–35

X. CONCLUSION

  • The survey provides a TL tutorial, reviews applications across wireless-network issues, and outlines challenges, open issues, and future research directions.Its application coverage includes spectrum management, localization, signal recognition, security, human activity recognition, and caching.Source: §X · p. 35

Related papers

Loading 2102.07572v2…