Since the proposal of transformers, these models have been limited to bounded input lengths, because of their need to attend to every token in the input. In this work, we propose Unlimiformer: a general approach that wraps any existing pretrained encoder-decoder transformer, and offloads the cross-attention computation to a single k-nearest-neighbor (kNN) index, while the returned kNN distances are the attention dot-product scores. This kNN index can be kept on either the GPU or CPU memory and queried in sub-linear time; this way, we can index practically unlimited input sequences, while every attention head in every decoder layer retrieves its top-k keys, instead of attending to every key. We evaluate Unlimiformer on several long-document and book-summarization benchmarks, showing that it can process even 500k token-long inputs from the BookSum dataset, without any input truncation at test time. We demonstrate that Unlimiformer improves pretrained models such as BART and Longformer by extending them to unlimited inputs without additional learned weights and without modifying their code. We make our code and models publicly available at https://github.com/abertsch72/unlimiformer .
Colonoscopy screening is the gold standard procedure for assessing abnormalities in the colon and rectum, such as ulcers and cancerous polyps. Measuring the abnormal mucosal area and its 3D reconstruction can help quantify the surveyed area and objectively evaluate disease burden. However, due to the complex topology of these organs and variable physical conditions, for example, lighting, large homogeneous texture, and image modality estimating distance from the camera aka depth) is highly challenging. Moreover, most colonoscopic video acquisition is monocular, making the depth estimation a non-trivial problem. While methods in computer vision for depth estimation have been proposed and advanced on natural scene datasets, the efficacy of these techniques has not been widely quantified on colonoscopy datasets. As the colonic mucosa has several low-texture regions that are not well pronounced, learning representations from an auxiliary task can improve salient feature extraction, allowing estimation of accurate camera depths. In this work, we propose to develop a novel multi-task learning (MTL) approach with a shared encoder and two decoders, namely a surface normal decoder and a depth estimator decoder. Our depth estimator incorporates attention mechanisms to enhance global context awareness. We leverage the surface normal prediction to improve geometric feature extraction. Also, we apply a cross-task consistency loss among the two geometrically related tasks, surface normal and camera depth. We demonstrate an improvement of 14.17% on relative error and 10.4% improvement on $\delta_{1}$ accuracy over the most accurate baseline state-of-the-art BTS approach. All experiments are conducted on a recently released C3VD dataset; thus, we provide a first benchmark of state-of-the-art methods.
Remarkable progress has been made in 3D reconstruction from single-view RGB-D inputs. MCC is the current state-of-the-art method in this field, which achieves unprecedented success by combining vision Transformers with large-scale training. However, we identified two key limitations of MCC: 1) The Transformer decoder is inefficient in handling large number of query points; 2) The 3D representation struggles to recover high-fidelity details. In this paper, we propose a new approach called NU-MCC that addresses these limitations. NU-MCC includes two key innovations: a Neighborhood decoder and a Repulsive Unsigned Distance Function (Repulsive UDF). First, our Neighborhood decoder introduces center points as an efficient proxy of input visual features, allowing each query point to only attend to a small neighborhood. This design not only results in much faster inference speed but also enables the exploitation of finer-scale visual features for improved recovery of 3D textures. Second, our Repulsive UDF is a novel alternative to the occupancy field used in MCC, significantly improving the quality of 3D object reconstruction. Compared to standard UDFs that suffer from holes in results, our proposed Repulsive UDF can achieve more complete surface reconstruction. Experimental results demonstrate that NU-MCC is able to learn a strong 3D representation, significantly advancing the state of the art in single-view 3D reconstruction. Particularly, it outperforms MCC by 9.7% in terms of the F1-score on the CO3D-v2 dataset with more than 5x faster running speed.
We present a novel alignment-before-generation approach to tackle the challenging task of generating general 3D shapes based on 2D images or texts. Directly learning a conditional generative model from images or texts to 3D shapes is prone to producing inconsistent results with the conditions because 3D shapes have an additional dimension whose distribution significantly differs from that of 2D images and texts. To bridge the domain gap among the three modalities and facilitate multi-modal-conditioned 3D shape generation, we explore representing 3D shapes in a shape-image-text-aligned space. Our framework comprises two models: a Shape-Image-Text-Aligned Variational Auto-Encoder (SITA-VAE) and a conditional Aligned Shape Latent Diffusion Model (ASLDM). The former model encodes the 3D shapes into the shape latent space aligned to the image and text and reconstructs the fine-grained 3D neural fields corresponding to given shape embeddings via the transformer-based decoder. The latter model learns a probabilistic mapping function from the image or text space to the latent shape space. Our extensive experiments demonstrate that our proposed approach can generate higher-quality and more diverse 3D shapes that better semantically conform to the visual or textural conditional inputs, validating the effectiveness of the shape-image-text-aligned space for cross-modality 3D shape generation.
Text-to-image diffusion models can create stunning images from natural language descriptions that rival the work of professional artists and photographers. However, these models are large, with complex network architectures and tens of denoising iterations, making them computationally expensive and slow to run. As a result, high-end GPUs and cloud-based inference are required to run diffusion models at scale. This is costly and has privacy implications, especially when user data is sent to a third party. To overcome these challenges, we present a generic approach that, for the first time, unlocks running text-to-image diffusion models on mobile devices in less than $2$ seconds. We achieve so by introducing efficient network architecture and improving step distillation. Specifically, we propose an efficient UNet by identifying the redundancy of the original model and reducing the computation of the image decoder via data distillation. Further, we enhance the step distillation by exploring training strategies and introducing regularization from classifier-free guidance. Our extensive experiments on MS-COCO show that our model with $8$ denoising steps achieves better FID and CLIP scores than Stable Diffusion v$1.5$ with $50$ steps. Our work democratizes content creation by bringing powerful text-to-image diffusion models to the hands of users.
Length generalization, the ability to generalize from small training context sizes to larger ones, is a critical challenge in the development of Transformer-based language models. Positional encoding (PE) has been identified as a major factor influencing length generalization, but the exact impact of different PE schemes on extrapolation in downstream tasks remains unclear. In this paper, we conduct a systematic empirical study comparing the length generalization performance of decoder-only Transformers with five different position encoding approaches including Absolute Position Embedding (APE), T5's Relative PE, ALiBi, and Rotary, in addition to Transformers without positional encoding (NoPE). Our evaluation encompasses a battery of reasoning and mathematical tasks. Our findings reveal that the most commonly used positional encoding methods, such as ALiBi, Rotary, and APE, are not well suited for length generalization in downstream tasks. More importantly, NoPE outperforms other explicit positional encoding methods while requiring no additional computation. We theoretically demonstrate that NoPE can represent both absolute and relative PEs, but when trained with SGD, it mostly resembles T5's relative PE attention patterns. Finally, we find that scratchpad is not always helpful to solve length generalization and its format highly impacts the model's performance. Overall, our work suggests that explicit position embeddings are not essential for decoder-only Transformers to generalize well to longer sequences.
We introduce SwiftSage, a novel agent framework inspired by the dual-process theory of human cognition, designed to excel in action planning for complex interactive reasoning tasks. SwiftSage integrates the strengths of behavior cloning and prompting large language models (LLMs) to enhance task completion performance. The framework comprises two primary modules: the Swift module, representing fast and intuitive thinking, and the Sage module, emulating deliberate thought processes. The Swift module is a small encoder-decoder LM fine-tuned on the oracle agent's action trajectories, while the Sage module employs LLMs such as GPT-4 for subgoal planning and grounding. We develop a heuristic method to harmoniously integrate the two modules, resulting in a more efficient and robust problem-solving process. In 30 tasks from the ScienceWorld benchmark, SwiftSage significantly outperforms other methods such as SayCan, ReAct, and Reflexion, demonstrating its effectiveness in solving complex interactive tasks.
Segmentation of road obstacles, including negative and positive obstacles, is critical to the safe navigation of autonomous vehicles. Recent methods have witnessed an increasing interest in using multi-modal data fusion (e.g., RGB and depth/disparity images). Although improved segmentation accuracy has been achieved by these methods, we still find that their performance could be easily degraded if the two modalities have inconsistent information, for example, distant obstacles that can be viewed in RGB images but cannot be viewed in depth/disparity images. To address this issue, we propose a novel two-encoder-two-decoder RGB-depth/disparity multi-modal network with Residual-Guided Fusion modules. Different from most existing networks that fuse feature maps in encoders, we fuse feature maps in decoder. We also release a large-scale RGB-depth/disparity dataset recorded in both urban and rural environments with manually-labeled ground truth for both negative- and positive-obstacles segmentation. Extensive experimental results demonstrate that our network achieves state-of-the-art performance compared with other networks.
Conjugate heat transfer (CHT) analyses are vital for the design of many energy systems. However, high-fidelity CHT numerical simulations are computationally intensive, which limits their applications such as design optimization, where hundreds to thousands of evaluations are required. In this work, we develop a modular deep encoder-decoder hierarchical (DeepEDH) convolutional neural network, a novel deep-learning-based surrogate modeling methodology for computationally intensive CHT analyses. Leveraging convective temperature dependencies, we propose a two-stage temperature prediction architecture that couples velocity and temperature fields. The proposed DeepEDH methodology is demonstrated by modeling the pressure, velocity, and temperature fields for a liquid-cooled cold-plate-based battery thermal management system with variable channel geometry. A computational mesh and CHT formulation of the cold plate is created and solved using the finite element method (FEM), generating a dataset of 1,500 simulations. Our performance analysis covers the impact of the novel architecture, separate DeepEDH models for each field, output geometry masks, multi-stage temperature field predictions, and optimizations of the hyperparameters and architecture. Furthermore, we quantify the influence of the CHT analysis' thermal boundary conditions on surrogate model performance, highlighting improved temperature model performance with higher heat fluxes. Compared to other deep learning neural network surrogate models, such as U-Net and DenseED, the proposed DeepEDH architecture for CHT analyses exhibits up to a 65% enhancement in the coefficient of determination $R^{2}$. (*Due to the notification of arXiv "The Abstract field cannot be longer than 1,920 characters", the appeared Abstract is shortened. For the full Abstract, please download the Article.)
Modern speech understanding (SU) runs a sophisticated pipeline: ingesting streaming voice input, the pipeline executes encoder-decoder based deep neural networks repeatedly; by doing so, the pipeline generates tentative outputs (called hypotheses), and periodically scores the hypotheses. This paper sets to accelerate SU on resource-constrained edge devices. It takes a hybrid approach: to speed up on-device execution; to offload inputs that are beyond the device's capacity. While the approach is well-known, we address SU's unique challenges with novel techniques: (1) late contextualization, which executes a model's attentive encoder in parallel to the input ingestion; (2) pilot inference, which mitigates the SU pipeline's temporal load imbalance; (3) autoregression offramps, which evaluate offloading decisions based on pilot inferences and hypotheses. Our techniques are compatible with existing speech models, pipelines, and frameworks; they can be applied independently or in combination. Our prototype, called PASU, is tested on Arm platforms with 6 - 8 cores: it delivers SOTA accuracy; it reduces the end-to-end latency by 2x and reduces the offloading needs by 2x.
Direct speech-to-speech translation (S2ST) translates speech from one language into another using a single model. However, due to the presence of linguistic and acoustic diversity, the target speech follows a complex multimodal distribution, posing challenges to achieving both high-quality translations and fast decoding speeds for S2ST models. In this paper, we propose DASpeech, a non-autoregressive direct S2ST model which realizes both fast and high-quality S2ST. To better capture the complex distribution of the target speech, DASpeech adopts the two-pass architecture to decompose the generation process into two steps, where a linguistic decoder first generates the target text, and an acoustic decoder then generates the target speech based on the hidden states of the linguistic decoder. Specifically, we use the decoder of DA-Transformer as the linguistic decoder, and use FastSpeech 2 as the acoustic decoder. DA-Transformer models translations with a directed acyclic graph (DAG). To consider all potential paths in the DAG during training, we calculate the expected hidden states for each target token via dynamic programming, and feed them into the acoustic decoder to predict the target mel-spectrogram. During inference, we select the most probable path and take hidden states on that path as input to the acoustic decoder. Experiments on the CVSS Fr-En benchmark demonstrate that DASpeech can achieve comparable or even better performance than the state-of-the-art S2ST model Translatotron 2, while preserving up to 18.53x speedup compared to the autoregressive baseline. Compared with the previous non-autoregressive S2ST model, DASpeech does not rely on knowledge distillation and iterative decoding, achieving significant improvements in both translation quality and decoding speed. Furthermore, DASpeech shows the ability to preserve the speaker's voice of the source speech during translation.
Neural combinatorial optimization (NCO) is a promising learning-based approach for solving challenging combinatorial optimization problems without specialized algorithm design by experts. However, most constructive NCO methods cannot solve problems with large-scale instance sizes, which significantly diminishes their usefulness for real-world applications. In this work, we propose a novel Light Encoder and Heavy Decoder (LEHD) model with a strong generalization ability to address this critical issue. The LEHD model can learn to dynamically capture the relationships between all available nodes of varying sizes, which is beneficial for model generalization to problems of various scales. Moreover, we develop a data-efficient training scheme and a flexible solution construction mechanism for the proposed LEHD model. By training on small-scale problem instances, the LEHD model can generate nearly optimal solutions for the Travelling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) with up to 1000 nodes, and also generalizes well to solve real-world TSPLib and CVRPLib problems. These results confirm our proposed LEHD model can significantly improve the state-of-the-art performance for constructive NCO. The code is available at https://github.com/CIAM-Group/NCO_code/tree/main/single_objective/LEHD.
As text generative models can give increasingly long answers, we tackle the problem of synthesizing long text in digital ink. We show that the commonly used models for this task fail to generalize to long-form data and how this problem can be solved by augmenting the training data, changing the model architecture and the inference procedure. These methods use contrastive learning technique and are tailored specifically for the handwriting domain. They can be applied to any encoder-decoder model that works with digital ink. We demonstrate that our method reduces the character error rate on long-form English data by half compared to baseline RNN and by 16% compared to the previous approach that aims at addressing the same problem. We show that all three parts of the method improve recognizability of generated inks. In addition, we evaluate synthesized data in a human study and find that people perceive most of generated data as real.
Joint punctuated and normalized automatic speech recognition (ASR), that outputs transcripts with and without punctuation and casing, remains challenging due to the lack of paired speech and punctuated text data in most ASR corpora. We propose two approaches to train an end-to-end joint punctuated and normalized ASR system using limited punctuated data. The first approach uses a language model to convert normalized training transcripts into punctuated transcripts. This achieves a better performance on out-of-domain test data, with up to 17% relative Punctuation-Case-aware Word Error Rate (PC-WER) reduction. The second approach uses a single decoder conditioned on the type of output. This yields a 42% relative PC-WER reduction compared to Whisper-base and a 4% relative (normalized) WER reduction compared to the normalized output of a punctuated-only model. Additionally, our proposed modeldemonstrates the feasibility of a joint ASR system using as little as 5% punctuated training data with a moderate (2.42% absolute) PC-WER increase.
The Manchu language, with its roots in the historical Manchurian region of Northeast China, is now facing a critical threat of extinction, as there are very few speakers left. In our efforts to safeguard the Manchu language, we introduce Mergen, the first-ever attempt at a Manchu-Korean Machine Translation (MT) model. To develop this model, we utilize valuable resources such as the Manwen Laodang(a historical book) and a Manchu-Korean dictionary. Due to the scarcity of a Manchu-Korean parallel dataset, we expand our data by employing word replacement guided by GloVe embeddings, trained on both monolingual and parallel texts. Our approach is built around an encoder-decoder neural machine translation model, incorporating a bi-directional Gated Recurrent Unit (GRU) layer. The experiments have yielded promising results, showcasing a significant enhancement in Manchu-Korean translation, with a remarkable 20-30 point increase in the BLEU score.
Cross-modal Steganography is the practice of concealing secret signals in publicly available cover signals (distinct from the modality of the secret signals) unobtrusively. While previous approaches primarily concentrated on concealing a relatively small amount of information, we propose THInImg, which manages to hide lengthy audio data (and subsequently decode talking head video) inside an identity image by leveraging the properties of human face, which can be effectively utilized for covert communication, transmission and copyright protection. THInImg consists of two parts: the encoder and decoder. Inside the encoder-decoder pipeline, we introduce a novel architecture that substantially increase the capacity of hiding audio in images. Moreover, our framework can be extended to iteratively hide multiple audio clips into an identity image, offering multiple levels of control over permissions. We conduct extensive experiments to prove the effectiveness of our method, demonstrating that THInImg can present up to 80 seconds of high quality talking-head video (including audio) in an identity image with 160x160 resolution.
With the development of Deep Neural Networks (DNNs), many efforts have been made to handle medical image segmentation. Traditional methods such as nnUNet train specific segmentation models on the individual datasets. Plenty of recent methods have been proposed to adapt the foundational Segment Anything Model (SAM) to medical image segmentation. However, they still focus on discrete representations to generate pixel-wise predictions, which are spatially inflexible and scale poorly to higher resolution. In contrast, implicit methods learn continuous representations for segmentation, which is crucial for medical image segmentation. In this paper, we propose I-MedSAM, which leverages the benefits of both continuous representations and SAM, to obtain better cross-domain ability and accurate boundary delineation. Since medical image segmentation needs to predict detailed segmentation boundaries, we designed a novel adapter to enhance the SAM features with high-frequency information during Parameter-Efficient Fine-Tuning (PEFT). To convert the SAM features and coordinates into continuous segmentation output, we utilize Implicit Neural Representation (INR) to learn an implicit segmentation decoder. We also propose an uncertainty-guided sampling strategy for efficient learning of INR. Extensive evaluations on 2D medical image segmentation tasks have shown that our proposed method with only 1.6M trainable parameters outperforms existing methods including discrete and implicit methods. The code will be available at: https://github.com/ucwxb/I-MedSAM.
Traditional 2D pose estimation models are limited by their category-specific design, making them suitable only for predefined object categories. This restriction becomes particularly challenging when dealing with novel objects due to the lack of relevant training data. To address this limitation, category-agnostic pose estimation (CAPE) was introduced. CAPE aims to enable keypoint localization for arbitrary object categories using a few-shot single model, requiring minimal support images with annotated keypoints. We present a significant departure from conventional CAPE techniques, which treat keypoints as isolated entities, by treating the input pose data as a graph. We leverage the inherent geometrical relations between keypoints through a graph-based network to break symmetry, preserve structure, and better handle occlusions. We validate our approach on the MP-100 benchmark, a comprehensive dataset comprising over 20,000 images spanning over 100 categories. Our solution boosts performance by 0.98% under a 1-shot setting, achieving a new state-of-the-art for CAPE. Additionally, we enhance the dataset with skeleton annotations. Our code and data are publicly available.
In this paper, we introduce U-Net v2, a new robust and efficient U-Net variant for medical image segmentation. It aims to augment the infusion of semantic information into low-level features while simultaneously refining high-level features with finer details. For an input image, we begin by extracting multi-level features with a deep neural network encoder. Next, we enhance the feature map of each level by infusing semantic information from higher-level features and integrating finer details from lower-level features through Hadamard product. Our novel skip connections empower features of all the levels with enriched semantic characteristics and intricate details. The improved features are subsequently transmitted to the decoder for further processing and segmentation. Our method can be seamlessly integrated into any Encoder-Decoder network. We evaluate our method on several public medical image segmentation datasets for skin lesion segmentation and polyp segmentation, and the experimental results demonstrate the segmentation accuracy of our new method over state-of-the-art methods, while preserving memory and computational efficiency. Code is available at: https://github.com/yaoppeng/U-Net_v2
Image matching that finding robust and accurate correspondences across images is a challenging task under extreme conditions. Capturing local and global features simultaneously is an important way to mitigate such an issue but recent transformer-based decoders were still stuck in the issues that CNN-based encoders only extract local features and the transformers lack locality. Inspired by the locality and implicit positional encoding of convolutions, a novel convolutional transformer is proposed to capture both local contexts and global structures more sufficiently for detector-free matching. Firstly, a universal FPN-like framework captures global structures in self-encoder as well as cross-decoder by transformers and compensates local contexts as well as implicit positional encoding by convolutions. Secondly, a novel convolutional transformer module explores multi-scale long range dependencies by a novel multi-scale attention and further aggregates local information inside dependencies for enhancing locality. Finally, a novel regression-based sub-pixel refinement module exploits the whole fine-grained window features for fine-level positional deviation regression. The proposed method achieves superior performances on a wide range of benchmarks. The code will be available on https://github.com/zwh0527/LGFCTR.