Xiaoyun Liu

Xiaoyun Liu, Divya Saxena, Jiannong Cao et al.

Neural architecture search (NAS) has gained significant traction in automating the design of neural networks. To reduce search time, differentiable architecture search (DAS) reframes the traditional paradigm of discrete candidate sampling and evaluation into a differentiable optimization over a super-net, followed by discretization. However, most existing DAS methods primarily focus on optimizing the coarse-grained operation-level topology, while neglecting finer-grained structures such as filter-level and weight-level patterns. This limits their ability to balance model performance with model size. Additionally, many methods compromise search quality to save memory during the search process. To tackle these issues, we propose Multi-Granularity Differentiable Architecture Search (MG-DARTS), a unified framework which aims to discover both effective and efficient architectures from scratch by comprehensively yet memory-efficiently exploring a multi-granularity search space. Specifically, we improve the existing DAS methods in two aspects. First, we adaptively adjust the retention ratios of searchable units across different granularity levels through adaptive pruning, which is achieved by learning granularity-specific discretization functions along with the evolving architecture. Second, we decompose the super-net optimization and discretization into multiple stages, each operating on a sub-net, and introduce progressive re-evaluation to enable re-pruning and regrowth of previous units, thereby mitigating potential bias. Extensive experiments on CIFAR-10, CIFAR-100 and ImageNet demonstrate that MG-DARTS outperforms other state-of-the-art methods in achieving a better trade-off between model accuracy and parameter efficiency. Codes are available at https://github.com/lxy12357/MG_DARTS.

Xiaoyun Liu, Divya Saxena, Jiannong Cao et al.

Structured pruning is widely used to compress large language models (LLMs), yet its effectiveness depends heavily on neuron importance estimation. Most existing methods estimate neuron importance from activation statistics on a single calibration dataset, which introduces calibration bias and degrades downstream cross-task generalization. We observe that neurons exhibit heterogeneous distribution sensitivity, with distribution-robust neurons maintaining consistent rankings across datasets and distribution-sensitive neurons showing high cross-dataset ranking variance. Based on this, we identify two structural limitations in existing methods. First, ranking all neurons within a shared space causes distribution-sensitive neurons that strongly activate on calibration inputs to dominate, crowding out distribution-robust neurons critical for out-of-distribution tasks. Second, applying activation-based importance metrics uniformly can be unreliable. Distribution-sensitive neurons that infrequently activate on calibration data receive insufficient activation signal for accurate local ranking. To address these limitations, we propose GPrune-LLM, a generalization-aware structured pruning framework that explicitly accounts for neuron differences in cross-distribution behavior. We first partition neurons into behavior-consistent modules to localize ranking competition, then evaluate activation-based metric reliability per module according to distribution sensitivity and score magnitude. For modules where activation-based scoring is unreliable, we switch to an activation-independent metric. Finally, we adaptively learn module-wise sparsity. Extensive experiments across multiple downstream tasks demonstrate GPrune-LLM's consistent improvements in post-compression generalization, particularly at high sparsity, and reduced dependence on importance metric choice.

Yuqing Zhao, Jiannong Cao, Divya Saxena et al.

In continual learning (CL), model growth enhances adaptability to new data. However, when model growth is applied improperly, especially in task-agnostic CL, where the entire grown model is used for inference, it can lead to severe degradation of learned knowledge, a problem we term growth-induced forgetting. Most existing methods that adopt model growth to improve adaptability often overlook the forgetting issue, resulting in compromised knowledge retention, making them unsuitable for task-agnostic settings. To promote both adaptability and knowledge retention with model growth, we identify the key: gradient and parameter sparsity. Introducing SparseGrow, which increases gradient sparsity through layer expansion and gradient gating to enable focused updates on parameters while preserving critical parameters, thus inhibiting forgetting. Moreover, it promotes parameter sparsity with sparse initialization and training, aiming at better control of model plasticity, improving adaptability over new data. Extensive experiments across diverse datasets, task-agnostic settings, and a large number of tasks demonstrate the necessity of controlled layer expansion and validate the effectiveness of SparseGrow in achieving high adaptability while minimizing forgetting in continual learning. By enabling model growth with sparsified gradients and parameters, SparseGrow paves the way for building scalable lifelong learning systems capable of continual adaptation with better knowledge retention.