Zheng Wang

Wei Wu, Qiuyi Li, Mingyang Li et al.

Advancements in DNA sequencing technologies have significantly improved our ability to decode genomic sequences. However, the prediction and interpretation of these sequences remain challenging due to the intricate nature of genetic material. Large language models (LLMs) have introduced new opportunities for biological sequence analysis. Recent developments in genomic language models have underscored the potential of LLMs in deciphering DNA sequences. Nonetheless, existing models often face limitations in robustness and application scope, primarily due to constraints in model structure and training data scale. To address these limitations, we present GENERator, a generative genomic foundation model featuring a context length of 98k base pairs (bp) and 1.2B parameters. Trained on an expansive dataset comprising 386B bp of eukaryotic DNA, the GENERator demonstrates state-of-the-art performance across both established and newly proposed benchmarks. The model adheres to the central dogma of molecular biology, accurately generating protein-coding sequences that translate into proteins structurally analogous to known families. It also shows significant promise in sequence optimization, particularly through the prompt-responsive generation of enhancer sequences with specific activity profiles. These capabilities position the GENERator as a pivotal tool for genomic research and biotechnological advancement, enhancing our ability to interpret and predict complex biological systems and enabling precise genomic interventions. Implementation details and supplementary resources are available at https://github.com/GenerTeam/GENERator.

Huawei Zhang, Chunwei Xia, Zheng Wang

Language models (LMs) underpin emerging mobile and embedded AI applications like meeting and video summarization and document analysis, which often require processing multiple long-context inputs. Running an LM locally on-device improves privacy, enables offline use, and reduces cost, but long-context inference quickly hits a \emph{memory capacity wall} as the key-value (KV) cache grows linearly with context length and batch size. We present KVSwap, a software framework to break this memory wall by offloading the KV cache to non-volatile secondary storage (disk). KVSwap leverages the observation that only a small, dynamically changing subset of KV entries is critical for generation. It stores the full cache on disk, uses a compact in-memory metadata to predict which entries to preload, overlaps computation with hardware-aware disk access, and orchestrates read patterns to match storage device characteristics. Our evaluation shows that across representative LMs and storage types, KVSwap delivers higher throughput under tight memory budgets while maintaining the generation quality when compared with existing KV cache offloading schemes.

Wei Wu, Zhuoshi Pan, Chao Wang et al.

Rapid advances in Large Language Models (LLMs) have spurred demand for processing extended context sequences in contemporary applications. However, this progress faces two challenges: performance degradation due to sequence lengths out-of-distribution, and excessively long inference times caused by the quadratic computational complexity of attention. These issues limit LLMs in long-context scenarios. In this paper, we propose Dynamic Token-Level KV Cache Selection (TokenSelect), a training-free method for efficient and accurate long-context inference. TokenSelect builds upon the observation of non-contiguous attention sparsity, using QK dot products to measure per-head KV Cache criticality at token-level. By per-head soft voting mechanism, TokenSelect selectively involves a few critical KV cache tokens in attention calculation without sacrificing accuracy. To further accelerate TokenSelect, we design the Selection Cache based on observations of consecutive Query similarity and implemented the efficient Paged Dot Product Kernel, significantly reducing the selection overhead. A comprehensive evaluation of TokenSelect demonstrates up to $23.84\times$ speedup in attention computation and up to $2.28\times$ acceleration in end-to-end latency, while providing superior performance compared to state-of-the-art long-context inference methods.

Di Hu, Zheng Wang, Haoyi Xiong et al.

Associating sound and its producer in complex audiovisual scene is a challenging task, especially when we are lack of annotated training data. In this paper, we present a flexible audiovisual model that introduces a soft-clustering module as the audio and visual content detector, and regards the pervasive property of audiovisual concurrency as the latent supervision for inferring the correlation among detected contents. To ease the difficulty of audiovisual learning, we propose a novel curriculum learning strategy that trains the model from simple to complex scene. We show that such ordered learning procedure rewards the model the merits of easy training and fast convergence. Meanwhile, our audiovisual model can also provide effective unimodal representation and cross-modal alignment performance. We further deploy the well-trained model into practical audiovisual sound localization and separation task. We show that our localization model significantly outperforms existing methods, based on which we show comparable performance in sound separation without referring external visual supervision. Our video demo can be found at https://youtu.be/kuClfGG0cFU.

Zheng Wang, Jianwu Li, Ge Song et al.

Self-attention (SA) mechanisms can capture effectively global dependencies in deep neural networks, and have been applied to natural language processing and image processing successfully. However, SA modules for image reconstruction have high time and space complexity, which restrict their applications to higher-resolution images. In this paper, we refine the SA module in self-attention generative adversarial networks (SAGAN) via adapting a non-local operation, revising the connectivity among the units in SA module and re-implementing its computational pattern, such that its time and space complexity is reduced from $\text{O}(n^2)$ to $\text{O}(n)$, but it is still equivalent to the original SA module. Further, we explore the principles behind the module and discover that our module is a special kind of channel attention mechanisms. Experimental results based on two benchmark datasets of image reconstruction, verify that under the same computational environment, two models can achieve comparable effectiveness for image reconstruction, but the proposed one runs faster and takes up less memory space.

Junjun Jiang, Yi Yu, Zheng Wang et al.

By developing sophisticated image priors or designing deep(er) architectures, a variety of image Super-Resolution (SR) approaches have been proposed recently and achieved very promising performance. A natural question that arises is whether these methods can be reformulated into a unifying framework and whether this framework assists in SR reconstruction? In this paper, we present a simple but effective single image SR method based on ensemble learning, which can produce a better performance than that could be obtained from any of SR methods to be ensembled (or called component super-resolvers). Based on the assumption that better component super-resolver should have larger ensemble weight when performing SR reconstruction, we present a Maximum A Posteriori (MAP) estimation framework for the inference of optimal ensemble weights. Specially, we introduce a reference dataset, which is composed of High-Resolution (HR) and Low-Resolution (LR) image pairs, to measure the super-resolution abilities (prior knowledge) of different component super-resolvers. To obtain the optimal ensemble weights, we propose to incorporate the reconstruction constraint, which states that the degenerated HR image should be equal to the LR observation one, as well as the prior knowledge of ensemble weights into the MAP estimation framework. Moreover, the proposed optimization problem can be solved by an analytical solution. We study the performance of the proposed method by comparing with different competitive approaches, including four state-of-the-art non-deep learning based methods, four latest deep learning based methods and one ensemble learning based method, and prove its effectiveness and superiority on three public datasets.