We introduce collaborative learning in which multiple classifier heads of the same network are simultaneously trained on the same training data to improve generalization and robustness to label noise with no extra inference cost. It acquires the strengths from auxiliary training, multi-task learning and knowledge distillation. There are two important mechanisms involved in collaborative learning. First, the consensus of multiple views from different classifier heads on the same example provides supplementary information as well as regularization to each classifier, thereby improving generalization. Second, intermediate-level representation (ILR) sharing with backpropagation rescaling aggregates the gradient flows from all heads, which not only reduces training computational complexity, but also facilitates supervision to the shared layers. The empirical results on CIFAR and ImageNet datasets demonstrate that deep neural networks learned as a group in a collaborative way significantly reduce the generalization error and increase the robustness to label noise.

In this paper, packet scheduling in an orthogonal frequency division multiplex access downlink is investigated based on cross-layer design and optimization. We first develop max-delay-utility scheduling with the help of channel and queue state information to exploit multiuser diversity and guarantee quality of service. The stability property of a scheduling policy is characterized by the stability region, which is the largest region on arrival rates for which the queueing system can be stabilized by the scheduling policy. It is shown in this paper that under very loose conditions, the max-delay-utility scheduling has the maximum stability region. In environments with insufficient scattering or strong line-of-sight components, delay transmit diversity can increase the fluctuation in the frequency domain to improve the performance. The simulation results show that the max-delay-utility scheduling with a frugality constraint is highly advantageous to data transmission with a low latency requirement over shared multiple channels, and that joint scheduling and power allocation can substantially boost the performance.

Next generation wireless systems, such as the IEEE 802.16e specification, are designed to provide great flexibility in scheduling radio resources on a per frame basis. Such flexibility, coupled with the need to maintain good quality of service (QoS) for cell-edge users can, however, lead to significant control channel overhead especially when using low code rates to maintain cell edge coverage. It is shown here that in interference-limited scenarios dual antenna mobile stations (MS) equipped with interference cancellation (IC) receivers can significantly alleviate this problem. In this study, the performance of a linear MMSE IC receiver is evaluated both at the link level and at the network level. It is shown that a practical IC receiver can achieve link performance gains in the range of 2.5-14 dB over a maximal ratio combining (MRC) receiver in single interferer dominant conditions. Performance at the network level is evaluated using the well-known EESM method to model the link-level performance for OFDM systems, and the introduction of IC receivers at the MS is shown to increase control channel coverage by 10-24% at low codeword repetition factors

The IEEE 802.16e system for multicast and broadcast service (MBS) is further investigated over [4]. It is shown that 802.16e networks can exploit advanced transmission and flexible resource allocation techniques to deliver high- performance MBS delivery within the constraints of the WiMAXRelease 1 system profile. A novel per-subcarrier SINR model for single frequency network (SFN) operation is proposed for the scenario in which cyclic shift transmit diversity (CSTD) - a form of delay diversity - is applied at the base station. The utility of CSTD for 802.16e MBS delivery in term of rate and coverage enhancement is investigated in detail based on the exponential effective-SINR mapping (EESM) for frame error rate (FER) prediction. In addition, an asymptotic effective SINR approach is developed for low- complexity approximation of the EESM FER mapping, and the proposed approximation is shown to be accurate for MBS simulation, while significantly reducing simulation effort.

In this paper, we provide an asymptotic performance analysis of channel-aware packet scheduling based on extreme value theory. We first address the average throughput of systems with a homogeneous average signal-to-noise ratio (SNR) and obtain its asymptotic expression. Compared with the exact throughput expression, the asymptotic one, which is applicable to a broader range of fading channels, is more concise and easier to get insights. Furthermore, we confirm the accuracy of the asymptotic results by theoretical analysis and numerical simulation. For a system with heterogeneous SNRs, normalized-SNR-based scheduling needs to be used for fairness. We also investigate the asymptotic average throughput of the normalized-SNR-based scheduling, and prove that the average throughput in this case is less than that in the homogeneous case with a power constraint.

In this paper, we provide an asymptotic performance analysis of channel-aware packet scheduling based on the extreme value theory. We first address the average throughput of systems with a homogeneous average signal-to-noise ratio (SNR), and obtain its asymptotic expression. Compared with the exact throughput expression, the asymptotic one, which is applicable to a broader range of fading channels, is more concise and easier from which to get insights. Furthermore, we confirm the accuracy of the asymptotic results by theoretical analysis and numerical simulation. For a system with heterogeneous SNRs, normalized-SNR-based scheduling needs to be used for fairness. We also investigate the asymptotic average throughput of the normalized-SNR-based scheduling, and prove that the average throughput in this case is less than that in the homogeneous case with a power constraint

In this paper, we provide an asymptotic performance analysis of channel-aware packet scheduling based on extreme value theory. We first address the average throughput of systems with a homogeneous average signal-to-noise ratio (SNR) and obtain its asymptotic expression. Compared to the exact throughput expression, the asymptotic one, which is applicable to a broader range of fading channels, is more concise and easier to get insights. The accuracy of the asymptotic results is confirmed by numerical simulation. Furthermore, we investigate the asymptotic average throughput of the normalized-SNR-based scheduling, which can provide fairness for a system with heterogeneous SNRs. The results show that the average throughput in this case is less than that in the homogeneous case with a power constraint.

In this paper, we highlight the advantage of multichannel random access schemes by jointly exploiting the inherent multiuser and multicarrier diversity. For a deeper understanding of network behavior, we define a parameter called the population density of the network and consider multichannel systems under different types of network density. We show that by exploiting the inherent diversity of the multicarrier networks, the asymptotic stable throughput (AST) of any single carrier medium access protocol can be increased. We have developed closed form expressions for the amount of gain that can be obtained using multicarrier diversity. We then take an example of the traditional Aloha systems to demonstrate the advantages of multicarrier systems over single carrier systems and show using simulation results that the maximum throughput gain achievable is as high as 50%.

Channel-aware packet scheduling is investigated in this paper. Based on extreme value theory, we propose asymptotic analyses for average throughput and delay in both single-carrier and multicarrier networks. Those analyses provide accurate results and have concise expressions, which explicitly show the multiuser diversity gain inherent in channel-aware scheduling. Although multicarrier networks with channel-aware scheduling have the same throughput as single-carrier networks, multicarrier networks can provide better delay performance than single-carrier networks

In this paper, we provide a theoretical framework for cross-layer optimization for orthogonal frequency division multiplexing (OFDM) wireless networks. The utility is used in our study to build a bridge between the physical layer and the media access control (MAC) layer and to balance the efficiency and fairness of wireless resource allocation. We formulate the cross-layer optimization problem as one that maximizes the average utility of all active users subject to certain conditions, which are determined by adaptive resource allocation schemes. We present necessary and sufficient conditions for utility-based optimal subcarrier assignment and power allocation and discuss the convergence properties of optimization. Numerical results demonstrate a significant performance gain for the cross-layer optimization and the gain increases with the number of active users in the networks.

We have established a theoretical framework for cross-layer optimization in orthogonal frequency division multiplexing (OFDM) wireless networks. In this paper, we focus on effective and practical algorithms for efficient and fair resource allocation in OFDM wireless networks. We have taken various conditions into account and developed a variety of efficient algorithms, including sorting-search dynamic subcarrier assignment, greedy bit loading, and power allocation, as well as objective aggregation algorithms. We have also modified those algorithms for a certain type of nonconcave utility functions. To further improve performance by exploiting time diversity, a low-pass time filter can be easily incorporated into all of the algorithms. Simulation results have confirmed that the utility-based cross-layer optimization can significantly enhance the system performance and guarantee fairness. The gains come from multiuser diversity, frequency diversity, as well as time diversity. The fairness is automatically achieved by the behavior of marginal utility functions.

This paper investigates multiuser downlink data scheduling with quality-of-service (QoS) provisioning over multiple shared fading channels, which, for a network point of view, provides line flexibility and granularity for resource allocation. A user-centric metric-a utility function with respect to mean waiting time-which is able to maintain fairness among users while providing delay QoS to individual users is used. This paper proposes scheduling algorithms that are aware of both channel and queue state information to achieve the maximum aggregate utility in the network. Simulation results confirm the significant performance and stability improvement provided by the utility-based scheduling scheme balancing multiuser diversity and queueing delay.

This paper investigates efficient and fair resource allocation, which is realized by using utility functions to measure the quality of service of users, in multiuser orthogonal frequency division multiplexing (OFDM) networks. To achieve the optimal resource allocation, we develop a series of efficient utility-based resource allocation algorithms that cover many conditions, including various dynamic subcarrier assignment and power allocation algorithms. The simulation results show that utility functions, together with dynamic subcarrier allocation, adaptive power allocation, and low-pass time window, can obtain multiuser, frequency, and time diversity, and significantly improve network efficiency as well as fairness.

In this paper, joint physical-MAC layer optimization in OFDM systems using adaptive modulation and subcarrier allocation is investigated based on utility theory. Since utility qualifies the level of users' satisfaction derived from the radio resources they occupy, it is ideal for a network optimization metric. We formulate the cross-layer optimization problem as one that maximizes the sum of the utilities over all active users through rate adaptation and dynamic subcarrier allocation with the limited radio resource and time-varying wireless channel constraints. Two effective subcarrier allocation algorithms with low complexity are proposed to solve the constrained optimization problem. Simulation results demonstrate the significant improvement of utility-based cross-layer optimization in OFDM.

The performance of power-controlled CDMA with smart antennas is investigated. In contrast to traditional analyses, by considering the interaction between beamforming and power control as well as the correlation of all transmitter powers caused by power control, this paper provides a theoretical framework for analyzing the performance of joint power control and beamforming. Two major typical smart antennas are compared, and it is shown that optimum beamforming is more effective in improving the capacity of CDMA systems than the spatially matched filter, especially in the situations of multiple quality of service requirements or non-uniform user distribution.