The paper "Spectral Co-Distillation for Personalized Federated Learning" by Zihan Chen, Howard H. Yang, Tony Q.S. Quek, and Kai Fong Ernest Chong introduces a novel approach to personalized federated learning (PFL) that leverages spectral information to capture the (dis-)similarity between generic and personalized models. The authors propose a spectral co-distillation framework, which includes two main components: spectral distillation for personalized model training and spectral co-distillation for generic model training. This framework aims to enhance the performance of both generic and personalized models by distilling knowledge from the generic model to the personalized models and vice versa. The paper also introduces a wait-free local training protocol, which eliminates idle waiting time during global communication, thereby reducing the total runtime of PFL. Extensive experiments on multiple datasets with heterogeneous data settings demonstrate the outperformance and efficacy of the proposed spectral co-distillation method and the wait-free training protocol. The results show that the proposed framework achieves superior generalizability and communication efficiency compared to existing PFL methods and conventional federated learning (FL) baselines. Key contributions of the paper include: 1. The first use of spectral distillation in PFL to capture the (dis-)similarity between generic and personalized models. 2. The introduction of a bi-directional knowledge distillation framework between generic and personalized models. 3. The development of a wait-free local training protocol to reduce the total PFL runtime. The authors discuss the limitations of their approach, such as the lack of handling stragglers and adversarial attacks, and the assumption of synchronized network connections, and suggest future work in these areas.The paper "Spectral Co-Distillation for Personalized Federated Learning" by Zihan Chen, Howard H. Yang, Tony Q.S. Quek, and Kai Fong Ernest Chong introduces a novel approach to personalized federated learning (PFL) that leverages spectral information to capture the (dis-)similarity between generic and personalized models. The authors propose a spectral co-distillation framework, which includes two main components: spectral distillation for personalized model training and spectral co-distillation for generic model training. This framework aims to enhance the performance of both generic and personalized models by distilling knowledge from the generic model to the personalized models and vice versa. The paper also introduces a wait-free local training protocol, which eliminates idle waiting time during global communication, thereby reducing the total runtime of PFL. Extensive experiments on multiple datasets with heterogeneous data settings demonstrate the outperformance and efficacy of the proposed spectral co-distillation method and the wait-free training protocol. The results show that the proposed framework achieves superior generalizability and communication efficiency compared to existing PFL methods and conventional federated learning (FL) baselines. Key contributions of the paper include: 1. The first use of spectral distillation in PFL to capture the (dis-)similarity between generic and personalized models. 2. The introduction of a bi-directional knowledge distillation framework between generic and personalized models. 3. The development of a wait-free local training protocol to reduce the total PFL runtime. The authors discuss the limitations of their approach, such as the lack of handling stragglers and adversarial attacks, and the assumption of synchronized network connections, and suggest future work in these areas.