Search

The storage and recall of factual associations in auto-regressive transformer language models (LMs) have drawn a great deal of attention, inspiring knowledge editing by directly modifying the located model weights. Most editing works achieve knowledge editing under the guidance of existing interpretations of knowledge recall that mainly focus on subject knowledge. However, these interpretations are seriously flawed, neglecting relation information and leading to the over-generalizing problem for editing. In this work, we discover a novel relation-focused perspective to interpret the knowledge recall of transformer LMs during inference and apply it on knowledge editing to avoid over-generalizing. Experimental results on the dataset supplemented with a new R-Specificity criterion demonstrate that our editing approach significantly alleviates over-generalizing while remaining competitive on other criteria, breaking the domination of subject-focused editing for future research.

As the volume and complexity increase, graph-structured data commonly need to be split and stored across distributed systems. To enable data mining on subgraphs within these distributed systems, federated graph learning has been proposed, allowing collaborative training of Graph Neural Networks (GNNs) across clients without sharing raw node features. However, when dealing with graph structures that involve high-order relationships between nodes, known as hypergraphs, existing federated graph learning methods are less effective. In this study, we introduce FedHGL, an innovative federated hypergraph learning algorithm. FedHGL is designed to collaboratively train a comprehensive hypergraph neural network across multiple clients, facilitating mining tasks on subgraphs of a hypergraph where relationships are not merely pairwise. To address the high-order information loss between subgraphs caused by distributed storage, we introduce a pre-propagation hyperedge completion operation before the federated training process. In this pre-propagation step, cross-client feature aggregation is performed and distributed at the central server to ensure that this information can be utilized by the clients. Furthermore, by incorporating local differential privacy (LDP) mechanisms, we ensure that the original node features are not disclosed during this aggregation process. Experimental results on seven real-world datasets confirm the effectiveness of our approach and demonstrate its performance advantages over traditional federated graph learning methods.

Parameter-efficient fine-tuning (PEFT) has emerged as the predominant technique for fine-tuning in the era of large language models. However, existing PEFT methods still have inadequate training efficiency. Firstly, the utilization of large-scale foundation models during the training process is excessively redundant for certain fine-tuning tasks. Secondly, as the model size increases, the growth in trainable parameters of empirically added PEFT modules becomes non-negligible and redundant, leading to inefficiency. To achieve task-specific efficient fine-tuning, we propose the Light-PEFT framework, which includes two methods: Masked Early Pruning of the Foundation Model and Multi-Granularity Early Pruning of PEFT. The Light-PEFT framework allows for the simultaneous estimation of redundant parameters in both the foundation model and PEFT modules during the early stage of training. These parameters can then be pruned for more efficient fine-tuning. We validate our approach on GLUE, SuperGLUE, QA tasks, and various models. With Light-PEFT, parameters of the foundation model can be pruned by up to over 40%, while still controlling trainable parameters to be only 25% of the original PEFT method. Compared to utilizing the PEFT method directly, Light-PEFT achieves training and inference speedup, reduces memory usage, and maintains comparable performance and the plug-and-play feature of PEFT., Comment: Findings of ACL 2024

Searching for exotic spin-dependent interactions that beyond the standard model has been of interest for past decades and is crucial for unraveling the mysteries of the universe. Previous laboratory searches primarily focus on searching for either static or modulated energy-level shifts caused by exotic spin-dependent interactions. Here, we introduce a theoretical model based on thermal motion of particles, providing another efficient way to search for exotic spin-dependent interactions. The theoretical model indicates that as the exotic spin-dependent interactions are related with the relative displacements and velocities of atoms, atoms undergoing thermal motion would experience a fluctuating energy-level shift induced by the exotic interactions. Moreover, the resulting exotic energy-level shift noise could be sensed by high-sensitivity instruments. By using the model and taking the high-sensitivity atomic magnetometer as an example, we set the most stringent laboratory experiment constraints on eight different kinds of exotic spin- and velocity-dependent interactions, with five of which at the force range below 1 cm have not been covered previously. Furthermore, this theoretical model can be easily applied in other fields of quantum sensing, such as atomic clocks, atom interferometers and NV-diamond sensors, to further improve the laboratory constraints on exotic spin-dependent interactions.