The recent Segment Anything Model (SAM) has demonstrated remarkable zero-shot capability and flexible geometric prompting in general image segmentation. However, SAM often struggles when handling various unconventional images, such as aerial, medical, and non-RGB images. This paper presents CAT-SAM, a ConditionAl Tuning network that adapts SAM toward various unconventional target tasks with just few-shot target samples. CAT-SAM freezes the entire SAM and adapts its mask decoder and image encoder simultaneously with a small number of learnable parameters. The core design is a prompt bridge structure that enables decoder-conditioned joint tuning of the heavyweight image encoder and the lightweight mask decoder. The bridging maps the prompt token of the mask decoder to the image encoder, fostering synergic adaptation of the encoder and the decoder with mutual benefits. We develop two representative tuning strategies for the image encoder which leads to two CAT-SAM variants: one injecting learnable prompt tokens in the input space and the other inserting lightweight adapter networks. Extensive experiments over 11 unconventional tasks show that both CAT-SAM variants achieve superior target segmentation performance consistently even under the very challenging one-shot adaptation setup. Project page: https://xiaoaoran.github.io/projects/CAT-SAM
In unsupervised medical image registration, the predominant approaches involve the utilization of a encoder-decoder network architecture, allowing for precise prediction of dense, full-resolution displacement fields from given paired images. Despite its widespread use in the literature, we argue for the necessity of making both the encoder and decoder learnable in such an architecture. For this, we propose a novel network architecture, termed LessNet in this paper, which contains only a learnable decoder, while entirely omitting the utilization of a learnable encoder. LessNet substitutes the learnable encoder with simple, handcrafted features, eliminating the need to learn (optimize) network parameters in the encoder altogether. Consequently, this leads to a compact, efficient, and decoder-only architecture for 3D medical image registration. Evaluated on two publicly available brain MRI datasets, we demonstrate that our decoder-only LessNet can effectively and efficiently learn both dense displacement and diffeomorphic deformation fields in 3D. Furthermore, our decoder-only LessNet can achieve comparable registration performance to state-of-the-art methods such as VoxelMorph and TransMorph, while requiring significantly fewer computational resources. Our code and pre-trained models are available at https://github.com/xi-jia/LessNet.
Recently, language models (LMs), especially large language models (LLMs), have revolutionized the field of deep learning. Both encoder-decoder models and prompt-based techniques have shown immense potential for natural language processing and code-based tasks. Over the past several years, many research labs and institutions have invested heavily in high-performance computing, approaching or breaching exascale performance levels. In this paper, we posit that adapting and utilizing such language model-based techniques for tasks in high-performance computing (HPC) would be very beneficial. This study presents our reasoning behind the aforementioned position and highlights how existing ideas can be improved and adapted for HPC tasks.
Pretrained Language Models (PLMs) have become the de facto starting point for fine-tuning on downstream tasks. However, as model sizes continue to increase, traditional fine-tuning of all the parameters becomes challenging. To address this, parameter-efficient fine-tuning (PEFT) methods have gained popularity as a means to adapt PLMs effectively. In parallel, recent studies have revealed the presence of activation sparsity within the intermediate outputs of the multilayer perceptron (MLP) blocks in transformers. Low activation density enables efficient model inference on sparsity-aware hardware. Building upon this insight, in this work, we propose a novel density loss that encourages higher activation sparsity (equivalently, lower activation density) in the pre-trained models. We demonstrate the effectiveness of our approach by utilizing mainstream PEFT techniques, including QLoRA, LoRA, Adapter, and Prompt/Prefix Tuning, to facilitate efficient model adaptation across diverse downstream tasks. Experiments show that our proposed method, \textbf{DEFT} (Density-Efficient Fine-Tuning), can consistently reduce activation density by up to \textbf{44.94\%} on RoBERTa$_\mathrm{Large}$ and by \textbf{53.19\%} (encoder density) and \textbf{90.60\%} (decoder density) on Flan-T5$_\mathrm{XXL}$ (\textbf{11B}) compared to PEFT, using GLUE and QA (SQuAD) benchmarks respectively. We also introduce \textbf{ADA-DEFT}, an adaptive variant of our DEFT approach, which achieves significant memory and runtime savings during inference. For instance, ADA-DEFT reduces runtime by \textbf{8.79\%}and memory usage by \textbf{17.46\%} in Flan-T5$_\mathrm{XL}$, and by \textbf{2.79\%} and \textbf{2.54\%} respectively in Flan-T5$_\mathrm{XXL}$. Additionally, we showcase that DEFT works complementarily with quantized and pruned models.
To help the open-source community have a better understanding of Mixture-of-Experts (MoE) based large language models (LLMs), we train and release OpenMoE, a series of fully open-sourced and reproducible decoder-only MoE LLMs, ranging from 650M to 34B parameters and trained on up to over 1T tokens. Our investigation confirms that MoE-based LLMs can offer a more favorable cost-effectiveness trade-off than dense LLMs, highlighting the potential effectiveness for future LLM development. One more important contribution of this study is an in-depth analysis of the routing mechanisms within our OpenMoE models, leading to three significant findings: Context-Independent Specialization, Early Routing Learning, and Drop-towards-the-End. We discovered that routing decisions in MoE models are predominantly based on token IDs, with minimal context relevance. The token-to-expert assignments are determined early in the pre-training phase and remain largely unchanged. This imperfect routing can result in performance degradation, particularly in sequential tasks like multi-turn conversations, where tokens appearing later in a sequence are more likely to be dropped. Finally, we rethink our design based on the above-mentioned observations and analysis. To facilitate future MoE LLM development, we propose potential strategies for mitigating the issues we found and further improving off-the-shelf MoE LLM designs.
We study incremental constituent parsers to assess their capacity to output trees based on prefix representations alone. Guided by strictly left-to-right generative language models and tree-decoding modules, we build parsers that adhere to a strong definition of incrementality across languages. This builds upon work that asserted incrementality, but that mostly only enforced it on either the encoder or the decoder. Finally, we conduct an analysis against non-incremental and partially incremental models.
In this paper, we propose the first framework that enables solving graph learning tasks of all levels (node, edge and graph) and all types (generation, regression and classification) using one formulation. We first formulate prediction tasks including regression and classification into a generic (conditional) generation framework, which enables diffusion models to perform deterministic tasks with provable guarantees. We then propose Latent Graph Diffusion (LGD), a generative model that can generate node, edge, and graph-level features of all categories simultaneously. We achieve this goal by embedding the graph structures and features into a latent space leveraging a powerful encoder and decoder, then training a diffusion model in the latent space. LGD is also capable of conditional generation through a specifically designed cross-attention mechanism. Leveraging LGD and the ``all tasks as generation'' formulation, our framework is capable of solving graph tasks of various levels and types. We verify the effectiveness of our framework with extensive experiments, where our models achieve state-of-the-art or highly competitive results across a wide range of generation and regression tasks.
Foundation models of time series have not been fully developed due to the limited availability of time series corpora and the underexploration of scalable pre-training. Based on the similar sequential formulation of time series and natural language, increasing research demonstrates the feasibility of leveraging large language models (LLM) for time series. Nevertheless, the inherent autoregressive property and decoder-only architecture of LLMs have not been fully considered, resulting in insufficient utilization of LLM abilities. To fully revitalize the general-purpose token transition and multi-step generation capability of large language models, we propose AutoTimes to repurpose LLMs as autoregressive time series forecasters, which projects time series into the embedding space of language tokens and autoregressively generates future predictions with arbitrary lengths. Compatible with any decoder-only LLMs, the consequent forecaster exhibits the flexibility of the lookback length and scalability with larger LLMs. Further, we formulate time series as prompts, extending the context for prediction beyond the lookback window, termed in-context forecasting. By introducing LLM-embedded textual timestamps, AutoTimes can utilize chronological information to align multivariate time series. Empirically, AutoTimes achieves state-of-the-art with 0.1% trainable parameters and over $5\times$ training/inference speedup compared to advanced LLM-based forecasters. Code is available at this repository: https://github.com/thuml/AutoTimes.
How to effectively interact audio with vision has garnered considerable interest within the multi-modality research field. Recently, a novel audio-visual segmentation (AVS) task has been proposed, aiming to segment the sounding objects in video frames under the guidance of audio cues. However, most existing AVS methods are hindered by a modality imbalance where the visual features tend to dominate those of the audio modality, due to a unidirectional and insufficient integration of audio cues. This imbalance skews the feature representation towards the visual aspect, impeding the learning of joint audio-visual representations and potentially causing segmentation inaccuracies. To address this issue, we propose AVSAC. Our approach features a Bidirectional Audio-Visual Decoder (BAVD) with integrated bidirectional bridges, enhancing audio cues and fostering continuous interplay between audio and visual modalities. This bidirectional interaction narrows the modality imbalance, facilitating more effective learning of integrated audio-visual representations. Additionally, we present a strategy for audio-visual frame-wise synchrony as fine-grained guidance of BAVD. This strategy enhances the share of auditory components in visual features, contributing to a more balanced audio-visual representation learning. Extensive experiments show that our method attains new benchmarks in AVS performance.
Image fusion integrates essential information from multiple images into a single composite, enhancing structures, textures, and refining imperfections. Existing methods predominantly focus on pixel-level and semantic visual features for recognition, but often overlook the deeper text-level semantic information beyond vision. Therefore, we introduce a novel fusion paradigm named image Fusion via vIsion-Language Model (FILM), for the first time, utilizing explicit textual information from source images to guide the fusion process. Specifically, FILM generates semantic prompts from images and inputs them into ChatGPT for comprehensive textual descriptions. These descriptions are fused within the textual domain and guide the visual information fusion, enhancing feature extraction and contextual understanding, directed by textual semantic information via cross-attention. FILM has shown promising results in four image fusion tasks: infrared-visible, medical, multi-exposure, and multi-focus image fusion. We also propose a vision-language dataset containing ChatGPT-generated paragraph descriptions for the eight image fusion datasets across four fusion tasks, facilitating future research in vision-language model-based image fusion. Code and dataset are available at https://github.com/Zhaozixiang1228/IF-FILM.
In the field of resource-constrained robots and the need for effective place recognition in multi-robotic systems, this article introduces RecNet, a novel approach that concurrently addresses both challenges. The core of RecNet's methodology involves a transformative process: it projects 3D point clouds into range images, compresses them using an encoder-decoder framework, and subsequently reconstructs the range image, restoring the original point cloud. Additionally, RecNet utilizes the latent vector extracted from this process for efficient place recognition tasks. This approach not only achieves comparable place recognition results but also maintains a compact representation, suitable for sharing among robots to reconstruct their collective maps. The evaluation of RecNet encompasses an array of metrics, including place recognition performance, the structural similarity of the reconstructed point clouds, and the bandwidth transmission advantages, derived from sharing only the latent vectors. Our proposed approach is assessed using both a publicly available dataset and field experiments$^1$, confirming its efficacy and potential for real-world applications.
Emerging Learned image Compression (LC) achieves significant improvements in coding efficiency by end-to-end training of neural networks for compression. An important benefit of this approach over traditional codecs is that any optimization criteria can be directly applied to the encoder-decoder networks during training. Perceptual optimization of LC to comply with the Human Visual System (HVS) is among such criteria, which has not been fully explored yet. This paper addresses this gap by proposing a novel framework to integrate Just Noticeable Distortion (JND) principles into LC. Leveraging existing JND datasets, three perceptual optimization methods are proposed to integrate JND into the LC training process: (1) Pixel-Wise JND Loss (PWL) prioritizes pixel-by-pixel fidelity in reproducing JND characteristics, (2) Image-Wise JND Loss (IWL) emphasizes on overall imperceptible degradation levels, and (3) Feature-Wise JND Loss (FWL) aligns the reconstructed image features with perceptually significant features. Experimental evaluations demonstrate the effectiveness of JND integration, highlighting improvements in rate-distortion performance and visual quality, compared to baseline methods. The proposed methods add no extra complexity after training.
Medical image segmentation using deep neural networks has been highly successful. However, the effectiveness of these networks is often limited by inadequate dense prediction and inability to extract robust features. To achieve refined dense prediction, we propose densely decoded networks (ddn), by selectively introducing 'crutch' network connections. Such 'crutch' connections in each upsampling stage of the network decoder (1) enhance target localization by incorporating high resolution features from the encoder, and (2) improve segmentation by facilitating multi-stage contextual information flow. Further, we present a training strategy based on adaptive deep supervision (ads), which exploits and adapts specific attributes of input dataset, for robust feature extraction. In particular, ads strategically locates and deploys auxiliary supervision, by matching the average input object size with the layer-wise effective receptive fields (lerf) of a network, resulting in a class of ddns. Such inclusion of 'companion objective' from a specific hidden layer, helps the model pay close attention to some distinct input-dependent features, which the network might otherwise 'ignore' during training. Our new networks and training strategy are validated on 4 diverse datasets of different modalities, demonstrating their effectiveness.
Can we model Non-Euclidean graphs as pure language or even Euclidean vectors while retaining their inherent information? The Non-Euclidean property have posed a long term challenge in graph modeling. Despite recent graph neural networks and graph transformers efforts encoding graphs as Euclidean vectors, recovering the original graph from vectors remains a challenge. In this paper, we introduce GraphsGPT, featuring an Graph2Seq encoder that transforms Non-Euclidean graphs into learnable Graph Words in the Euclidean space, along with a GraphGPT decoder that reconstructs the original graph from Graph Words to ensure information equivalence. We pretrain GraphsGPT on $100$M molecules and yield some interesting findings: (1) The pretrained Graph2Seq excels in graph representation learning, achieving state-of-the-art results on $8/9$ graph classification and regression tasks. (2) The pretrained GraphGPT serves as a strong graph generator, demonstrated by its strong ability to perform both few-shot and conditional graph generation. (3) Graph2Seq+GraphGPT enables effective graph mixup in the Euclidean space, overcoming previously known Non-Euclidean challenges. (4) The edge-centric pretraining framework GraphsGPT demonstrates its efficacy in graph domain tasks, excelling in both representation and generation. Code is available at \href{https://github.com/A4Bio/GraphsGPT}{GitHub}.
Recent advancements in self-supervised learning have unlocked the potential to harness unlabeled data for auxiliary tasks, facilitating the learning of beneficial priors. This has been particularly advantageous in fields like medical image analysis, where labeled data are scarce. Although effective for classification tasks, this methodology has shown limitations in more complex applications, such as medical image segmentation. In this paper, we introduce Medical imaging Enhanced with Dynamic Self-Adaptive Semantic Segmentation (MedSASS), a dedicated self-supervised framework tailored for medical image segmentation. We evaluate MedSASS against existing state-of-the-art methods across four diverse medical datasets, showcasing its superiority. MedSASS outperforms existing CNN-based self-supervised methods by 3.83% and matches the performance of ViT-based methods. Furthermore, when MedSASS is trained end-to-end, covering both encoder and decoder, it demonstrates significant improvements of 14.4% for CNNs and 6% for ViT-based architectures compared to existing state-of-the-art self-supervised strategies.
Most recent scribble-supervised segmentation methods commonly adopt a CNN framework with an encoder-decoder architecture. Despite its multiple benefits, this framework generally can only capture small-range feature dependency for the convolutional layer with the local receptive field, which makes it difficult to learn global shape information from the limited information provided by scribble annotations. To address this issue, this paper proposes a new CNN-Transformer hybrid solution for scribble-supervised medical image segmentation called ScribFormer. The proposed ScribFormer model has a triple-branch structure, i.e., the hybrid of a CNN branch, a Transformer branch, and an attention-guided class activation map (ACAM) branch. Specifically, the CNN branch collaborates with the Transformer branch to fuse the local features learned from CNN with the global representations obtained from Transformer, which can effectively overcome limitations of existing scribble-supervised segmentation methods. Furthermore, the ACAM branch assists in unifying the shallow convolution features and the deep convolution features to improve model's performance further. Extensive experiments on two public datasets and one private dataset show that our ScribFormer has superior performance over the state-of-the-art scribble-supervised segmentation methods, and achieves even better results than the fully-supervised segmentation methods. The code is released at https://github.com/HUANGLIZI/ScribFormer.
In the era of AI at the edge, self-driving cars, and climate change, the need for energy-efficient, small, embedded AI is growing. Spiking Neural Networks (SNNs) are a promising approach to address this challenge, with their event-driven information flow and sparse activations. We propose Spiking CenterNet for object detection on event data. It combines an SNN CenterNet adaptation with an efficient M2U-Net-based decoder. Our model significantly outperforms comparable previous work on Prophesee's challenging GEN1 Automotive Detection Dataset while using less than half the energy. Distilling the knowledge of a non-spiking teacher into our SNN further increases performance. To the best of our knowledge, our work is the first approach that takes advantage of knowledge distillation in the field of spiking object detection.
Deep neural networks have exhibited remarkable performance in a variety of computer vision fields, especially in semantic segmentation tasks. Their success is often attributed to multi-level feature fusion, which enables them to understand both global and local information from an image. However, we found that multi-level features from parallel branches are on different scales. The scale disequilibrium is a universal and unwanted flaw that leads to detrimental gradient descent, thereby degrading performance in semantic segmentation. We discover that scale disequilibrium is caused by bilinear upsampling, which is supported by both theoretical and empirical evidence. Based on this observation, we propose injecting scale equalizers to achieve scale equilibrium across multi-level features after bilinear upsampling. Our proposed scale equalizers are easy to implement, applicable to any architecture, hyperparameter-free, implementable without requiring extra computational cost, and guarantee scale equilibrium for any dataset. Experiments showed that adopting scale equalizers consistently improved the mIoU index across various target datasets, including ADE20K, PASCAL VOC 2012, and Cityscapes, as well as various decoder choices, including UPerHead, PSPHead, ASPPHead, SepASPPHead, and FCNHead.
The (variational) graph auto-encoder is widely used to learn representations for graph-structured data. However, the formation of real-world graphs is a complicated and heterogeneous process influenced by latent factors. Existing encoders are fundamentally holistic, neglecting the entanglement of latent factors. This reduces the effectiveness of graph analysis tasks, while also making it more difficult to explain the learned representations. As a result, learning disentangled graph representations with the (variational) graph auto-encoder poses significant challenges and remains largely unexplored in the current research. In this paper, we introduce the Disentangled Graph Auto-Encoder (DGA) and the Disentangled Variational Graph Auto-Encoder (DVGA) to learn disentangled representations. Specifically, we first design a disentangled graph convolutional network with multi-channel message-passing layers to serve as the encoder. This allows each channel to aggregate information about each latent factor. The disentangled variational graph auto-encoder's expressive capability is then enhanced by applying a component-wise flow to each channel. In addition, we construct a factor-wise decoder that takes into account the characteristics of disentangled representations. We improve the independence of representations by imposing independence constraints on the mapping channels for distinct latent factors. Empirical experiments on both synthetic and real-world datasets demonstrate the superiority of our proposed method compared to several state-of-the-art baselines.
For analysing real-world networks, graph representation learning is a popular tool. These methods, such as a graph autoencoder (GAE), typically rely on low-dimensional representations, also called embeddings, which are obtained through minimising a loss function; these embeddings are used with a decoder for downstream tasks such as node classification and edge prediction. While GAEs tend to be fairly accurate, they suffer from scalability issues. For improved speed, a Local2Global approach, which combines graph patch embeddings based on eigenvector synchronisation, was shown to be fast and achieve good accuracy. Here we propose L2G2G, a Local2Global method which improves GAE accuracy without sacrificing scalability. This improvement is achieved by dynamically synchronising the latent node representations, while training the GAEs. It also benefits from the decoder computing an only local patch loss. Hence, aligning the local embeddings in each epoch utilises more information from the graph than a single post-training alignment does, while maintaining scalability. We illustrate on synthetic benchmarks, as well as real-world examples, that L2G2G achieves higher accuracy than the standard Local2Global approach and scales efficiently on the larger data sets. We find that for large and dense networks, it even outperforms the slow, but assumed more accurate, GAEs.