Zihan Qiu

Zihan Qiu, Zhen Liu, Shuicheng Yan et al.

It has been shown that semi-parametric methods, which combine standard neural networks with non-parametric components such as external memory modules and data retrieval, are particularly helpful in data scarcity and out-of-distribution (OOD) scenarios. However, existing semi-parametric methods mostly depend on independent raw data points - this strategy is difficult to scale up due to both high computational costs and the incapacity of current attention mechanisms with a large number of tokens. In this paper, we introduce a novel heterogeneous memory augmentation approach for neural networks which, by introducing learnable memory tokens with attention mechanism, can effectively boost performance without huge computational overhead. Our general-purpose method can be seamlessly combined with various backbones (MLP, CNN, GNN, and Transformer) in a plug-and-play manner. We extensively evaluate our approach on various image and graph-based tasks under both in-distribution (ID) and OOD conditions and show its competitive performance against task-specific state-of-the-art methods. Code is available at \url{https://github.com/qiuzh20/HMA}.

Zihan Qiu, Zeyu Huang, Kaiyue Wen et al.

We investigate the functional role of emergent outliers in large language models, specifically attention sinks (a few tokens that consistently receive large attention logits) and residual sinks (a few fixed dimensions with persistently large activations across most tokens). We hypothesize that these outliers, in conjunction with the corresponding normalizations (\textit{e.g.}, softmax attention and RMSNorm), effectively rescale other non-outlier components. We term this phenomenon \textit{outlier-driven rescaling} and validate this hypothesis across different model architectures and training token counts. This view unifies the origin and mitigation of both sink types. Our main conclusions and observations include: (1) Outliers function jointly with normalization: removing normalization eliminates the corresponding outliers but degrades training stability and performance; directly clipping outliers while retaining normalization leads to degradation, indicating that outlier-driven rescaling contributes to training stability. (2) Outliers serve more as rescale factors rather than contributors, as the final contributions of attention and residual sinks are significantly smaller than those of non-outliers. (3) Outliers can be absorbed into learnable parameters or mitigated via explicit gated rescaling, leading to improved training performance (average gain of 2 points) and enhanced quantization robustness (1.2 points degradation under W4A4 quantization).

Hao Zhao, Zihan Qiu, Huijia Wu et al.

The Mixture of Experts (MoE) for language models has been proven effective in augmenting the capacity of models by dynamically routing each input token to a specific subset of experts for processing. Despite the success, most existing methods face a challenge for balance between sparsity and the availability of expert knowledge: enhancing performance through increased use of expert knowledge often results in diminishing sparsity during expert selection. To mitigate this contradiction, we propose HyperMoE, a novel MoE framework built upon Hypernetworks. This framework integrates the computational processes of MoE with the concept of knowledge transferring in multi-task learning. Specific modules generated based on the information of unselected experts serve as supplementary information, which allows the knowledge of experts not selected to be used while maintaining selection sparsity. Our comprehensive empirical evaluations across multiple datasets and backbones establish that HyperMoE significantly outperforms existing MoE methods under identical conditions concerning the number of experts.