Hao Dai

Continual Generalized Category Discovery (C-GCD) faces a critical challenge: incrementally learning new classes from unlabeled data streams while preserving knowledge of old classes. Existing methods struggle with catastrophic forgetting, especially when unlabeled data mixes known and novel categories. We address this by analyzing C-GCD’s forgetting dynamics through a Bayesian lens, revealing that covariance misalignment between old and new classes drives performance degradation. Building on this insight, we propose Variational Bayes C-GCD (VB-CGCD), a novel framework that integrates variational inference with covariance-aware nearest-class-mean classification. VB-CGCD adaptively aligns class distributions while suppressing pseudo-label noise via stochastic variational updates. Experiments show VB-CGCD surpasses prior art by +15.21% with the overall accuracy in the final session on standard benchmarks. We also introduce a new challenging benchmark with only 10% labeled data and extended online phases—VB-CGCD achieves a 67.86% final accuracy, significantly higher than state-of-the-art (38.55%), demonstrating its robust applicability across diverse scenarios.

With the fast growth of mobile edge computing (MEC), the deep neural network (DNN) has gained more opportunities in application to various mobile services. Given the tremendous number of learning parameters and large model size, the DNN model is often trained in cloud center and then dispatched to end devices for inference via edge network. Therefore, maximizing the cost-efficiency of learned model dispatch in the edge network would be a critical problem for the model serving in various application contexts. To reach this goal, in this article we focus mainly on reducing the total model dispatch cost in the edge network while maintaining the efficiency of the model inference. We first study this problem in its off-line form as a baseline where a sequence of n requests can be pre-defined in advance and exploit dynamic programming techniques to obtain a fast optimal algorithm in time complexity of O(m2n) under a semi-homogeneous cost model in a m -sized network. Then, we design and implement a 2.5-competitive algorithm for its online case with a provable lower bound of 2 for any deterministic online algorithm. We verify our results through careful algorithmic analysis and validate their actual performance via a trace-based study based on a public open international mobile network dataset.

Currently, most edge-based Deep Reinforcement Learning (Deep-RL) applications have been deployed in the edge network, however, their mainstream studies are still short of adequate considerations on its limited compute and bandwidth resources. In this paper, we investigate the near on-policy of actions taking in distributed Deep-RL architecture, and propose a “hybrid near on-policy” Deep-RL framework, called Coknight, by leveraging a game-theory based DNN partition approach. We first formulate the partition problem into a variant of knapsack problem in device-edge setting, and then transform it into a potential game with a formal proof. Finally, we show the problem is NP-complete whereby an efficient distributed algorithm based on the potential game theory is developed from device perspective to achieve fast and dynamic partitioning. Coknight not only significantly improves the resource efficiency of the Deep-RL but also allows the inference to enforce the scalability of the actor policy. We prototype the framework with extensive experiments to validate our findings. The experimental results show that with the premise of a rapid convergence guarantee, Coknight, compared with Seed-RL, can reduce GPU utilization by 30% while providing large-scale scalability.