Abstract Conditional Generative Adversarial Networks (cGANs) provide a flexible framework for learning conditional distributions, but they suffer from two fundamental challenges: the scarcity of conditional samples and the curse of dimensionality in the image space. In this work, we address both issues by introducing Vicinal Estimation (VE) into the cGAN framework and analyzing it through Barron-space discriminators. VE alleviates the lack of conditional samples by coupling nearby labels via an auxiliary sampling distribution, effectively transforming the problem into a collection of unconditional GANs in the vicinal label space. Meanwhile, the Barron-space analysis yields a dimension-independent generalization bound that holds irrespective of the image dimension, and we show how this bound transfers from VE conditionals back to the original conditional distributions. We develop VE-cGAN , a practical instantiation of this idea, and demonstrate through experiments on benchmark datasets that it achieves improved perceptual quality and label consistency compared with baselines. Our theoretical and empirical findings together highlight VE as a principled and effective approach to overcoming the lack of conditional samples and the curse of dimensionality in conditional generative modeling.

Over the past decade, there is a growing interest in collaborative learning that can enhance AI models of multiple parties. However, it is still challenging to enhance performance them without sharing private data and models from individual parties. One recent promising approach is to develop distillation-based algorithms that exploit unlabeled public data but the results are still unsatisfactory in both theory and practice. To tackle this problem, we rigorously analyze a representative distillation-based algorithm in the view of kernel regression. This work provides the first theoretical results to prove the (nearly) minimax optimality of the nonparametric collaborative learning algorithm that does not directly share local data or models in massively distributed statistically heterogeneous environments. Inspired by our theoretical results, we also propose a practical distillation-based collaborative learning algorithm based on neural network architecture. Our algorithm successfully bridges the gap between our theoretical assumptions and practical settings with neural networks through feature kernel matching. We simulate various regression tasks to verify our theory and demonstrate the practical feasibility of our proposed algorithm.

Federated Learning (FL) is a collaborative machine learning paradigm for data privacy preservation. Recently, a knowledge distillation (KD) based information sharing approach in FL, which conducts ensemble distillation on an unlabeled public dataset, has been proposed. However, despite its experimental success and usefulness, the theoretical analysis of the KD based approach has not been satisfactorily conducted. In this work, we build a theoretical foundation of the ensemble distillation framework in federated learning from the perspective of kernel ridge regression (KRR). In this end, we propose a KD based FL algorithm for KRR models which is related with some existing KD based FL algorithms, and analyze our algorithm theoretically. We show that our algorithm makes local prediction models as much powerful as the centralized KRR model (which is a KRR model trained by all of local datasets) in terms of the convergence rate of the generalization error if the unlabeled public dataset is sufficiently large. We also provide experimental results to verify our theoretical results on ensemble distillation in federated learning.

A Multimedia multipath (MMMP) system aims at offering higher reliability and stability utilizing a variety of communication media and paths. To satisfy quality of service (QoS) requirements, an associated router should estimate the path status and allocate traffic load to each path properly. In this paper, we propose an adaptive load balancing algorithm which does not require any information about the system and the necessary information is estimated through online learning with Gaussian Process Regression (GPR). To this end, we introduce a probing period to collect training data on delay for GPR and estimate the timeout probability of each path using the predictive distributions from GPR. We then analyze and minimize the cost function, the weighted sum of the timeout probabilities of the paths. In the analysis, we propose two approximation methods of the timeout probability without direct calculations, which allows the proposed algorithm to operate in online manner. Through extensive simulations under various scenarios with real-world traffic traces, we demonstrate that the proposed algorithm balances the traffic load properly according to dynamic system conditions. Furthermore, we analyze the impact of probing packets on the network and design the proposed algorithm to reduce the communication overhead due to probing packets.

Traffic prediction facilitates intelligent networking maintenance by enabling efficient network resource allocation. With the development of machine learning algorithms, traffic prediction has attracted increasing attention and has been widely used in resource allocation and traffic management. In this paper, we consider an input traffic with a quality of service (QoS) requirement, such as the overflow probability, and propose an adaptive bandwidth allocation method based on the Gaussian process regression (GPR) to satisfy the required QoS. In the proposed method with GPR, we consider the stochastic property of each sample path individually and compute the required bandwidth adaptively for each sample path by estimating its overflow probability. Thus, it is more beneficial than the bandwidth allocation method based on the average overflow probability over all sample paths that are widely used in many previous works. We investigate the computational complexity and performance of the proposed method through simulation with real-world traffic as well as computer-generated traffic and show that the proposed method allocates the bandwidth adaptively and efficiently to satisfy the required QoS.

We propose a distance-based inter-cell interference coordination (ICIC) scheme in small cell networks (SCNs). While most of the previous works focus on a randomly selected user called a typical user, we focus on an edge user because the main purpose of ICIC is to improve the performance of an edge user. Since there are many inactive small cell base stations (SBSs) in SCNs, a simple criterion for an edge user such as being located near a cell boundary is not appropriate for SCNs. To accurately detect edge users experiencing severe performance degradation, we newly define an edge user in SCNs based on the nearest active neighbor SBS. We then apply our scheme only to edge users where SBSs within so-called the cooperation radius from each edge user cooperate. With the help of the stochastic geometry we obtain a semi-closed expression for the coverage probability of an edge user with our scheme. We investigate two tradeoffs on the resource efficiency of a network and the coverage probability of an interior user. We then determine the optimal cooperation radius that maximizes the coverage probability of an edge user considering the two tradeoffs. Our analytical results are validated through simulations.

In this paper we propose a distance-based Inter-Cell Interference Coordination (ICIC) scheme for Small Cell Networks (SCNs). While most of the previous works focus on a typical user, not an edge user, and do not consider the spatial distribution of the edge users, we carefully consider the spatial distribution of edge users and then apply a distance-based ICIC scheme only to the edge users, where small cell base stations within a certain radius from each edge user cooperate to improve the coverage probability of edge users. With the help of the stochastic geometry theory we analyze the performance of the proposed ICIC scheme and derive the optimal ICIC cooperation radius.

Efficient sharing of the network spectrum is required to meet the exponentially growing demand of wireless communications, and the role of the MAC protocols has been becoming more important. In this regard, a novel MAC protocol called the Renewal Access Protocol (RAP) was recently proposed. It was shown that the RAP achieves optimal throughput, high short-term fairness, and near-optimal delay performances when the number of nodes in the network is known to all nodes. However, it is not easy for a node to know the number of nodes in the network and the number of nodes is time varying. So we propose an adaptive version of the RAP called the Adaptive Renewal Access Protocol (A-RAP) in this paper. A node equipped with the A-RAP estimates the number of nodes utilizing two dimensional Backoff Selection State (BSS), the Estimated Number of Nodes (ENN) and phase, based on its transmission results. For an efficient design of the A-RAP, we analyze the performance of the A-RAP and carefully determine two key parameters, the adjustment probabilities and the numbers of phases. Numerical and simulation results are provided to verify that the proposed A-RAP achieves good throughput and short-term fairness as the RAP. The results also show that the A-RAP achieves better short-term fairness than other MAC protocols proposed in the open literature.

As a promising solution for handling super dense wireless networks, wireless local area networks (WLANs) have been intensively considered due to their wide availability. However, the contention-based MAC protocol in WLANs should be modified because of its inefficiency. To this end, we consider a recently proposed novel MAC protocol called the renewal access protocol (RAP). With the RAP, we analyze two strategies for resolving collisions efficiently and achieving optimal throughput performance in super dense WLANs: strategies without and with grouping. First, we analyze the asymptotic behavior of the RAP itself (i.e., without grouping) as the number of terminals goes to infinity. We show that the RAP can achieve optimal throughput even in super dense WLANs and the relevant optimal access probability can be derived in a closed form. Second, we propose a grouping strategy in the RAP called the grouped RAP (G-RAP). While a grouping strategy in the IEEE 802.11ah standard is based on time division, our G-RAP is based on transmission attempts. So the G-RAP does not waste channel resources. We show that the G-RAP achieves the optimal network throughput for any group structure if terminals use the optimal access probability that we derive. Our analytical results are validated by simulation.

We consider a simple MAC protocol, called the renewal access protocol (RAP), that adopts all of the legacy 802.11 standard but the backoff stage feature. To meet two objectives in the design of the RAP—optimal throughput and high short-term fairness—we develop a mathematical model of the RAP and rigorously analyze the performance of the RAP. First, we show that the throughput performance of the RAP depends only on the expectation of the selection distribution where the backoff counter is selected, provided that the number of terminals is fixed, which is in accordance with a well-known result. Second, with the help of renewal and reliability theories, we analyze the short-term fairness of the RAP. We also show that if the RAP has a selection distribution of the New Better than Used in Expectation (NBUE) type, the RAP can guarantee high short-term fairness. Third, we construct a special binomial distribution that is obviously of the NBUE type that can achieve high short-term fairness as well as optimal throughput when used as the selection distribution of the RAP. Furthermore, by the Poisson approximation for binomial distributions, we propose to use in practice a Poisson distribution corresponding to the special binomial distribution. Numerical and simulation results are provided to validate our analysis.

This work seeks to develop an analytical model for the per-node throughput analysis of IEEE 802.11 WLAN networks with hidden nodes by extending the Bianchi's model. With the analytic model we derive the per-node throughput of each node and quantify the impact of hidden nodes on per-node throughput. Through our analysis, we find that nodes having more hidden nodes are likely to have worse throughput performance than nodes having less hidden nodes, so resulting in unfairness in per-node throughput.

We consider a distributed random channel access policy in a cognitive radio network where multiple secondary users (SUs) contend for spectrum usage over available multiple primary user (PU) channels. In the random channel access policy each SU stochastically determines whether to access a channel based on the access probability (AP) which is adapted to the sensing results. It is then important to obtain the optimal AP values from the SUs' throughput performance perspective. In our analysis, we assume that all SUs acquire a common channel state information that is potentially erroneous. We analyze the throughput of an arbitrary SU and the packet collision probability of a PU due to the interference by SUs. By considering a requirement on the packet collision probability of a PU, we rigorously derive an explicit expression on the optimal AP values that maximize the throughput of an SU under the requirement for PUs. In the derivation we show that the multi-dimensional optimization problem with AP values can be converted into a one-dimensional optimization problem, so that the optimal AP values can be easily found. Numerical examples are provided to validate our analysis and to investigate the performance behavior of the optimal channel access policy.

OFDMA is the basis of future broadband access, due to its many inherent advantages such as scalability and fine granularity for multi-user access. OFDMA-Aloha combines the flexibility of OFDMA with basic Aloha's collision resolution mechanism over sub-carriers, in an attempt to reduce packet collisions and achieve faster retransmission. However, this comes at the expense of a larger slot size, due to lower channel rates per subcarrier. The above gives rise to a fundamental question: whether to use a single wide-band Aloha channel and retransmit via random back-off in next K time slots, or to retransmit immediately in one of K narrow-band sub-channels which are each 1/K slower (OFDMA-Aloha)? We answer this question, by analyzing the two protocols: Aloha and OFDMA-Aloha under the same total bandwidth and load conditions. We first derive the exact distribution of the packet access delay of OFDMA-Aloha in the saturated case. Then, we extend the analysis to the unsaturated case and derive the mean queue length and packet delay by decomposing the system of interfering queues into multiple independent queues utilizing the symmetry in our system. Our results show that if the network is already saturated, channelization does not bring substantial reduction in the collision rate to the point where it outweighs the effect of expanded slot size. In this case the single channel Aloha performs better than OFDMA-Aloha especially when the gap between the number of channels and the number users is large. On other hand, when the network is lightly loaded, OFDMA-Aloha enjoys smaller packet delays, but not for long as it saturates faster than the single channel Aloha.