Hongjian Fang

Ziwei Xiang, Fanhu Zeng, Hongjian Fang et al.

Large Vision Language Models (LVLMs) have achieved remarkable success in a range of downstream tasks that require multimodal interaction, but their capabilities come with substantial computational and memory overhead, which hinders practical deployment. Among numerous acceleration techniques, post-training quantization is a popular and effective strategy for reducing memory cost and accelerating inference. However, existing LVLM quantization methods typically measure token sensitivity at the modality level, which fails to capture the complex cross-token interactions and falls short in quantitatively measuring the quantization error at the token level. As tokens interact within the model, the distinction between modalities gradually diminishes, suggesting the need for fine-grained calibration. Inspired by axiomatic attribution in mechanistic interpretability, we introduce a fine-grained quantization strategy on Quantization-aware Integrated Gradients (QIG), which leverages integrated gradients to quantitatively evaluate token sensitivity and push the granularity from modality level to token level, reflecting both inter-modality and intra-modality dynamics. Extensive experiments on multiple LVLMs under both W4A8 and W3A16 settings show that our method improves accuracy across models and benchmarks with negligible latency overhead. For example, under 3-bit weight-only quantization, our method improves the average accuracy of LLaVA-onevision-7B by 1.60%, reducing the gap to its full-precision counterpart to only 1.33%. The code is available at https://github.com/ucas-xiang/QIG.

Hongjian Fang, Yi Zeng, Jianbo Tang et al.

How neural networks in the human brain represent commonsense knowledge, and complete related reasoning tasks is an important research topic in neuroscience, cognitive science, psychology, and artificial intelligence. Although the traditional artificial neural network using fixed-length vectors to represent symbols has gained good performance in some specific tasks, it is still a black box that lacks interpretability, far from how humans perceive the world. Inspired by the grandmother-cell hypothesis in neuroscience, this work investigates how population encoding and spiking timing-dependent plasticity (STDP) mechanisms can be integrated into the learning of spiking neural networks, and how a population of neurons can represent a symbol via guiding the completion of sequential firing between different neuron populations. The neuron populations of different communities together constitute the entire commonsense knowledge graph, forming a giant graph spiking neural network. Moreover, we introduced the Reward-modulated spiking timing-dependent plasticity (R-STDP) mechanism to simulate the biological reinforcement learning process and completed the related reasoning tasks accordingly, achieving comparable accuracy and faster convergence speed than the graph convolutional artificial neural networks. For the fields of neuroscience and cognitive science, the work in this paper provided the foundation of computational modeling for further exploration of the way the human brain represents commonsense knowledge. For the field of artificial intelligence, this paper indicated the exploration direction for realizing a more robust and interpretable neural network by constructing a commonsense knowledge representation and reasoning spiking neural networks with solid biological plausibility.

Yao Liang, Hongjian Fang, Yi Zeng et al.

Reasoning and question answering as a basic cognitive function for humans, is nevertheless a great challenge for current artificial intelligence. Although the Differentiable Neural Computer (DNC) model could solve such problems to a certain extent, the development is still limited by its high algorithm complexity, slow convergence speed, and poor test robustness. Inspired by the learning and memory mechanism of the brain, this paper proposed a Memory Transformation based Differentiable Neural Computer (MT-DNC) model. MT-DNC incorporates working memory and long-term memory into DNC, and realizes the autonomous transformation of acquired experience between working memory and long-term memory, thereby helping to effectively extract acquired knowledge to improve reasoning ability. Experimental results on bAbI question answering task demonstrated that our proposed method achieves superior performance and faster convergence speed compared to other existing DNN and DNC models. Ablation studies also indicated that the memory transformation from working memory to long-term memory plays essential role in improving the robustness and stability of reasoning. This work explores how brain-inspired memory transformation can be integrated and applied to complex intelligent dialogue and reasoning systems.

Meng Li, Peisong Wang, Yuantian Shao et al.

Large Language Models (LLMs) achieve strong performance across diverse tasks but face deployment challenges due to their massive size. Structured pruning offers acceleration benefits but leads to significant performance degradation. Recent PCA-based pruning methods have alleviated this issue by retaining key activation components, but are only applied between modules in order to fuse the transformation matrix, which introduces extra parameters and severely disrupts activation distributions due to residual connections. To address these issues, we propose IntraSlice, a framework that applies block-wise module-intra PCA compression pruning. By leveraging the structural characteristics of Transformer modules, we design an approximate PCA method whose transformation matrices can be fully fused into the model without additional parameters. We also introduce a PCA-based global pruning ratio estimator that further considers the distribution of compressed activations, building on conventional module importance. We validate our method on Llama2, Llama3, and Phi series across various language benchmarks. Experimental results demonstrate that our approach achieves superior compression performance compared to recent baselines at the same compression ratio or inference speed.

Yuxi Li, Lu Si, Yujie Hou et al.

Advancements in remote sensing (RS) imagery have provided high-resolution detail and vast coverage, yet existing methods, such as image-level captioning/retrieval and object-level detection/segmentation, often fail to capture mid-scale semantic entities essential for interpreting large-scale scenes. To address this, we propose the conteXtual referring Map (XeMap) task, which focuses on contextual, fine-grained localization of text-referred regions in large-scale RS scenes. Unlike traditional approaches, XeMap enables precise mapping of mid-scale semantic entities that are often overlooked in image-level or object-level methods. To achieve this, we introduce XeMap-Network, a novel architecture designed to handle the complexities of pixel-level cross-modal contextual referring mapping in RS. The network includes a fusion layer that applies self- and cross-attention mechanisms to enhance the interaction between text and image embeddings. Furthermore, we propose a Hierarchical Multi-Scale Semantic Alignment (HMSA) module that aligns multiscale visual features with the text semantic vector, enabling precise multimodal matching across large-scale RS imagery. To support XeMap task, we provide a novel, annotated dataset, XeMap-set, specifically tailored for this task, overcoming the lack of XeMap datasets in RS imagery. XeMap-Network is evaluated in a zero-shot setting against state-of-the-art methods, demonstrating superior performance. This highlights its effectiveness in accurately mapping referring regions and providing valuable insights for interpreting large-scale RS environments.