Yumiao Zhao

Yumiao Zhao, Bo Jiang, Beibei Wang et al.

Visual Prompting (VP) has emerged as an efficient paradigm for adapting large-scale pre-trained vision models to downstream tasks by incorporating learnable prompts at the input level. However, existing VP methods typically employ dense pixel-level prompts, which often suffer from redundant perturbations, limited generalization and energy inefficiency. To overcome these limitations, we propose to integrate brain-inspired spiking learning into visual prompt learning tasks. As we know that spiking neuron can perform inexpensive information processing by transmitting the input data into discrete spike trains and return sparse outputs. Inspired by this, we propose \textbf{Lo}w-\textbf{R}ank visual \textbf{S}pike \textbf{P}rompting (LoRSP), a novel framework that learns dynamic low-rank sparse visual prompts naturally via a Spiking neuron learning mechanism. The core idea of LoRSP is to exploit the brain-inspired sparse firing mechanism of spiking neurons to generate pixel-level sparse prompt for each instance. To be specific, we first construct a series of prompt factors via low-rank factorization to capture distinct prompt subspaces. These prompt factors are then fed into an SNN architecture, which performs the integrate-and-fire process to emit spikes. As a result, our LoRSP generates a \emph{sparse} visual prompt while maintaining the low-rank constraint. This design enables instance-specific selective prompting, leading to more compact and robust adaptation across diverse downstream tasks. Extensive experiments on five heterogeneous vision backbones and multiple benchmarks demonstrate that LoRSP achieves competitive performance while requiring fewer tunable parameters compared to existing VP methods.

Zhexiang Zhang, Ye Wang, Yumiao Zhao et al.

Serving large Mixture-of-Experts (MoE) models is challenging because of their large memory footprints, heterogeneous resource demands, and highly dynamic inference workloads. Most existing MoE inference systems deploy the entire model as a monolithic unit, forcing attention and MoE layers to share the same resource configuration despite their different scaling behaviors and resource bottlenecks. Such coarse-grained provisioning leads to resource inefficiency and suboptimal performance. We present JANUS, a scalable and resource-efficient MoE inference system built around three key principles. First, JANUS disaggregates attention and MoE layers onto separate GPU worker pools, enabling independent resource provisioning for the two layer types, and uses an adaptive two-phase communication mechanism for low-latency data exchange. Second, because MoE-layer execution is often memory-bound and highly sensitive to activated-expert imbalance, JANUS introduces a lightweight, microsecond-scale activation scheduler that balances per-layer activated experts across MoE instances to reduce inference latency. Third, JANUS employs a fine-grained, SLO-aware resource scaling scheme that jointly selects attention resources, MoE resources, and expert placement to minimize GPU cost under token-level SLOs. Evaluation shows that JANUS improves per-GPU throughput by up to 4.7x over state-of-the-art MoE inference baselines while satisfying token-level latency SLOs.

Yumiao Zhao, Bo Jiang, Yuhe Ding et al.

Adapter-based approaches have garnered attention for fine-tuning pre-trained Vision-Language Models (VLMs) on few-shot classification tasks. These methods strive to develop a lightweight module that better aligns visual and (category) textual representations, thereby enhancing performance on downstream few-shot learning tasks. However, existing adapters generally learn/align (category) textual-visual modalities via explicit spatial proximity in the underlying embedding space, which i) fails to capture the inherent one-to-many associations between categories and image samples and ii) struggles to establish accurate associations between the unknown categories and images. To address these issues, inspired by recent works on hyperbolic learning, we develop a novel Latent Hierarchical Adapter (LatHAdapter) for fine-tuning VLMs on downstream few-shot classification tasks. The core of LatHAdapter is to exploit the latent semantic hierarchy of downstream training data and employ it to provide richer, fine-grained guidance for the adapter learning process. Specifically, LatHAdapter first introduces some learnable `attribute' prompts as the bridge to align categories and images. Then, it projects the categories, attribute prompts, and images within each batch in a hyperbolic space, and employs hierarchical regularization to learn the latent semantic hierarchy of them, thereby fully modeling the inherent one-to-many associations among categories, learnable attributes, and image samples. Extensive experiments on four challenging few-shot tasks show that the proposed LatHAdapter consistently outperforms many other fine-tuning approaches, particularly in adapting known classes and generalizing to unknown classes.