YuGuang Wang

Mutian Yang, Zisen Zhan, Yutong Chen et al.

Catastrophic forgetting during knowledge injection severely undermines the continual learning capability of large language models (LLMs). Although existing methods attempt to mitigate this issue, they often lack a foundational theoretical explanation. We establish a gradient-based theoretical framework to explain catastrophic forgetting. We first prove that strongly negative gradient similarity is a fundamental cause of forgetting. We then use gradient similarity to identify two types of neurons: conflicting neurons that induce forgetting and account for 50%-75% of neurons, and collaborative neurons that mitigate forgetting and account for 25%-50%. Based on this analysis, we propose a knowledge injection method, Collaborative Neural Learning (CNL). By freezing conflicting neurons and updating only collaborative neurons, CNL theoretically eliminates catastrophic forgetting under an infinitesimal learning rate eta and an exactly known mastered set. Experiments on five LLMs, four datasets, and four optimizers show that CNL achieves zero forgetting in in-set settings and reduces forgetting by 59.1%-81.7% in out-of-set settings.

Bingxin Zhou, Xinliang Liu, Yuehua Liu et al.

Many real-world relational systems, such as social networks and biological systems, contain dynamic interactions. When learning dynamic graph representation, it is essential to employ sequential temporal information and geometric structure. Mainstream work achieves topological embedding via message passing networks (e.g., GCN, GAT). The temporal evolution, on the other hand, is conventionally expressed via memory units (e.g., LSTM or GRU) that possess convenient information filtration in a gate mechanism. Though, such a design prevents large-scale input sequence due to the over-complicated encoding. This work learns from the philosophy of self-attention and proposes an efficient spectral-based neural unit that employs informative long-range temporal interaction. The developed spectral window unit (SWINIT) model predicts scalable dynamic graphs with assured efficiency. The architecture is assembled with a few simple effective computational blocks that constitute randomized SVD, MLP, and graph Framelet convolution. The SVD plus MLP module encodes the long-short-term feature evolution of the dynamic graph events. A fast framelet graph transform in the framelet convolution embeds the structural dynamics. Both strategies enhance the model's ability on scalable analysis. In particular, the iterative SVD approximation shrinks the computational complexity of attention to O(Nd\log(d)) for the dynamic graph with N edges and d edge features, and the multiscale transform of framelet convolution allows sufficient scalability in the network training. Our SWINIT achieves state-of-the-art performance on a variety of online continuous-time dynamic graph learning tasks, while compared to baseline methods, the number of its learnable parameters reduces by up to seven times.