Transformer-based NLP models are powerful but have high computational costs that limit deployment. Finetuned encoder-decoder models are popular in specialized domains and can outperform larger more generalized decoder-only models, such as GPT-4. We introduce a new configuration for encoder-decoder models that improves efficiency on structured output and decomposable tasks where multiple outputs are required for a single shared input. Our method, prompt-in-decoder (PiD), encodes the input once and decodes the output in parallel, boosting both training and inference efficiency by avoiding duplicate input encoding and increasing the operational intensity (ratio of numbers of arithmetic operation to memory access) of decoding process by sharing the input key-value cache. We achieve computation reduction that roughly scales with the number of subtasks, gaining up to 4.6x speed-up over state-of-the-art models for dialogue state tracking, summarization, and question-answering tasks, with comparable or better performance.
Despite the impressive performance of deep learning models across diverse tasks, their complexity poses challenges for interpretation. This challenge is particularly evident for audio signals, where conveying interpretations becomes inherently difficult. To address this issue, we introduce Listenable Maps for Audio Classifiers (L-MAC), a posthoc interpretation method that generates faithful and listenable interpretations. L-MAC utilizes a decoder on top of a pretrained classifier to generate binary masks that highlight relevant portions of the input audio. We train the decoder with a loss function that maximizes the confidence of the classifier decision on the masked-in portion of the audio while minimizing the probability of model output for the masked-out portion. Quantitative evaluations on both in-domain and out-of-domain data demonstrate that L-MAC consistently produces more faithful interpretations than several gradient and masking-based methodologies. Furthermore, a user study confirms that, on average, users prefer the interpretations generated by the proposed technique.
We introduce SceneScript, a method that directly produces full scene models as a sequence of structured language commands using an autoregressive, token-based approach. Our proposed scene representation is inspired by recent successes in transformers & LLMs, and departs from more traditional methods which commonly describe scenes as meshes, voxel grids, point clouds or radiance fields. Our method infers the set of structured language commands directly from encoded visual data using a scene language encoder-decoder architecture. To train SceneScript, we generate and release a large-scale synthetic dataset called Aria Synthetic Environments consisting of 100k high-quality in-door scenes, with photorealistic and ground-truth annotated renders of egocentric scene walkthroughs. Our method gives state-of-the art results in architectural layout estimation, and competitive results in 3D object detection. Lastly, we explore an advantage for SceneScript, which is the ability to readily adapt to new commands via simple additions to the structured language, which we illustrate for tasks such as coarse 3D object part reconstruction.
Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary classes given a few support images annotated with keypoints. Existing methods only rely on the features extracted at support keypoints to predict or refine the keypoints on query image, but a few support feature vectors are local and inadequate for CAPE. Considering that human can quickly perceive potential keypoints of arbitrary objects, we propose a novel framework for CAPE based on such potential keypoints (named as meta-points). Specifically, we maintain learnable embeddings to capture inherent information of various keypoints, which interact with image feature maps to produce meta-points without any support. The produced meta-points could serve as meaningful potential keypoints for CAPE. Due to the inevitable gap between inherency and annotation, we finally utilize the identities and details offered by support keypoints to assign and refine meta-points to desired keypoints in query image. In addition, we propose a progressive deformable point decoder and a slacked regression loss for better prediction and supervision. Our novel framework not only reveals the inherency of keypoints but also outperforms existing methods of CAPE. Comprehensive experiments and in-depth studies on large-scale MP-100 dataset demonstrate the effectiveness of our framework.
Foundation models have reshaped the landscape of Remote Sensing (RS) by enhancing various image interpretation tasks. Pretraining is an active research topic, encompassing supervised and self-supervised learning methods to initialize model weights effectively. However, transferring the pretrained models to downstream tasks may encounter task discrepancy due to their formulation of pretraining as image classification or object discrimination tasks. In this study, we explore the Multi-Task Pretraining (MTP) paradigm for RS foundation models to address this issue. Using a shared encoder and task-specific decoder architecture, we conduct multi-task supervised pretraining on the SAMRS dataset, encompassing semantic segmentation, instance segmentation, and rotated object detection. MTP supports both convolutional neural networks and vision transformer foundation models with over 300 million parameters. The pretrained models are finetuned on various RS downstream tasks, such as scene classification, horizontal and rotated object detection, semantic segmentation, and change detection. Extensive experiments across 14 datasets demonstrate the superiority of our models over existing ones of similar size and their competitive performance compared to larger state-of-the-art models, thus validating the effectiveness of MTP.
In this paper, we present Paramanu-Ayn, a collection of legal language models trained exclusively on Indian legal case documents. This 97-million-parameter Auto-Regressive (AR) decoder-only model was pretrained from scratch with a context size of 8192 on a single GPU for just 185 hours, achieving an efficient MFU of 41.35. We also developed a legal domain specialized BPE tokenizer. We evaluated our model using perplexity and zero-shot tasks: case judgment prediction with explanation and abstractive case summarization. Paramanu-Ayn outperformed Llama-2 7B and Gemini-Pro in case judgment prediction with explanation task on test accuracy by nearly 2 percentage points, despite being 72 times smaller. In zero-shot abstractive summarization, it surpassed decoder-only LLMs generating fixed-length summaries (5000 tokens) by over 10 percentage points in BLEU and METEOR metrics, and by nearly 4 percentage points in BERTScore. Further evaluations on zero-shot commonsense and mathematical benchmarks showed that Paramanu-Ayn excelled despite being trained exclusively on legal documents, outperforming Llama-1, Llama-2, and Falcon on AGIEVAL-AQuA-RAT and AGIEVAL-SAT-Math tasks. We also instruction-tuned our model on 10,763 diverse legal tasks, including legal clause generation, legal drafting, case summarization, etc. The Paramanu-Ayn-instruct model scored above 8 out of 10 in clarity, relevance, completeness, and legal reasoning metrics by GPT-3.5-Turbo. We found that our models, were able to learn drafting knowledge and generalize to draft legal contracts and legal clauses with limited instruction-tuning. Hence, we conclude that for a strong domain-specialized generative language model (such as legal), domain specialized pretraining from scratch is more cost effective, environmentally friendly, and remains competitive with larger models or even better than adapting LLMs for legal domain tasks.
Modern Neural Machine Translation systems exhibit strong performance in several different languages and are constantly improving. Their ability to learn continuously is, however, still severely limited by the catastrophic forgetting issue. In this work, we leverage a key property of encoder-decoder Transformers, i.e. their generative ability, to propose a novel approach to continually learning Neural Machine Translation systems. We show how this can effectively learn on a stream of experiences comprising different languages, by leveraging a replay memory populated by using the model itself as a generator of parallel sentences. We empirically demonstrate that our approach can counteract catastrophic forgetting without requiring explicit memorization of training data. Code will be publicly available upon publication. Code: https://github.com/m-resta/sg-rep
There have been emerging research interest and advances in speech-to-speech translation (S2ST), translating utterances from one language to another. This work proposes Multitask Speech Language Model (MSLM), which is a decoder-only speech language model trained in a multitask setting. Without reliance on text training data, our model is able to support multilingual S2ST with speaker style preserved.
The progress in natural language processing (NLP) using large language models (LLMs) has greatly improved patient information extraction from clinical narratives. However, most methods based on the fine-tuning strategy have limited transfer learning ability for cross-domain applications. This study proposed a novel approach that employs a soft prompt-based learning architecture, which introduces trainable prompts to guide LLMs toward desired outputs. We examined two types of LLM architectures, including encoder-only GatorTron and decoder-only GatorTronGPT, and evaluated their performance for the extraction of social determinants of health (SDoH) using a cross-institution dataset from the 2022 n2c2 challenge and a cross-disease dataset from the University of Florida (UF) Health. The results show that decoder-only LLMs with prompt tuning achieved better performance in cross-domain applications. GatorTronGPT achieved the best F1 scores for both datasets, outperforming traditional fine-tuned GatorTron by 8.9% and 21.8% in a cross-institution setting, and 5.5% and 14.5% in a cross-disease setting.
In this work, we introduce FaceXFormer, an end-to-end unified transformer model capable of performing nine facial analysis tasks including face parsing, landmark detection, head pose estimation, attribute prediction, and estimation of age, gender, race, expression, and face visibility within a single framework. Conventional methods in face analysis have often relied on task-specific designs and pre-processing techniques, which limit their scalability and integration into a unified architecture. Unlike these conventional methods, FaceXFormer leverages a transformer-based encoder-decoder architecture where each task is treated as a learnable token, enabling the seamless integration and simultaneous processing of multiple tasks within a single framework. Moreover, we propose a novel parameter-efficient decoder, FaceX, which jointly processes face and task tokens, thereby learning generalized and robust face representations across different tasks. We jointly trained FaceXFormer on nine face perception datasets and conducted experiments against specialized and multi-task models in both intra-dataset and cross-dataset evaluations across multiple benchmarks, showcasing state-of-the-art or competitive performance. Further, we performed a comprehensive analysis of different backbones for unified face task processing and evaluated our model "in-the-wild", demonstrating its robustness and generalizability. To the best of our knowledge, this is the first work to propose a single model capable of handling nine facial analysis tasks while maintaining real-time performance at 33.21 FPS.
In this paper, we propose a novel video depth estimation approach, FutureDepth, which enables the model to implicitly leverage multi-frame and motion cues to improve depth estimation by making it learn to predict the future at training. More specifically, we propose a future prediction network, F-Net, which takes the features of multiple consecutive frames and is trained to predict multi-frame features one time step ahead iteratively. In this way, F-Net learns the underlying motion and correspondence information, and we incorporate its features into the depth decoding process. Additionally, to enrich the learning of multiframe correspondence cues, we further leverage a reconstruction network, R-Net, which is trained via adaptively masked auto-encoding of multiframe feature volumes. At inference time, both F-Net and R-Net are used to produce queries to work with the depth decoder, as well as a final refinement network. Through extensive experiments on several benchmarks, i.e., NYUDv2, KITTI, DDAD, and Sintel, which cover indoor, driving, and open-domain scenarios, we show that FutureDepth significantly improves upon baseline models, outperforms existing video depth estimation methods, and sets new state-of-the-art (SOTA) accuracy. Furthermore, FutureDepth is more efficient than existing SOTA video depth estimation models and has similar latencies when comparing to monocular models
Understanding illumination and reducing the need for supervision pose a significant challenge in low-light enhancement. Current approaches are highly sensitive to data usage during training and illumination-specific hyper-parameters, limiting their ability to handle unseen scenarios. In this paper, we propose a new zero-reference low-light enhancement framework trainable solely with normal light images. To accomplish this, we devise an illumination-invariant prior inspired by the theory of physical light transfer. This prior serves as the bridge between normal and low-light images. Then, we develop a prior-to-image framework trained without low-light data. During testing, this framework is able to restore our illumination-invariant prior back to images, automatically achieving low-light enhancement. Within this framework, we leverage a pretrained generative diffusion model for model ability, introduce a bypass decoder to handle detail distortion, as well as offer a lightweight version for practicality. Extensive experiments demonstrate our framework's superiority in various scenarios as well as good interpretability, robustness, and efficiency. Code is available on our project homepage: http://daooshee.github.io/QuadPrior-Website/
Despite significant advancements in image customization with diffusion models, current methods still have several limitations: 1) unintended changes in non-target areas when regenerating the entire image; 2) guidance solely by a reference image or text descriptions; and 3) time-consuming fine-tuning, which limits their practical application. In response, we introduce a tuning-free framework for simultaneous text-image-guided image customization, enabling precise editing of specific image regions within seconds. Our approach preserves the semantic features of the reference image subject while allowing modification of detailed attributes based on text descriptions. To achieve this, we propose an innovative attention blending strategy that blends self-attention features in the UNet decoder during the denoising process. To our knowledge, this is the first tuning-free method that concurrently utilizes text and image guidance for image customization in specific regions. Our approach outperforms previous methods in both human and quantitative evaluations, providing an efficient solution for various practical applications, such as image synthesis, design, and creative photography.
Open-vocabulary video instance segmentation strives to segment and track instances belonging to an open set of categories in a videos. The vision-language model Contrastive Language-Image Pre-training (CLIP) has shown robust zero-shot classification ability in image-level open-vocabulary tasks. In this paper, we propose a simple encoder-decoder network, called CLIP-VIS, to adapt CLIP for open-vocabulary video instance segmentation. Our CLIP-VIS adopts frozen CLIP and introduces three modules, including class-agnostic mask generation, temporal topK-enhanced matching, and weighted open-vocabulary classification. Given a set of initial queries, class-agnostic mask generation introduces a pixel decoder and a transformer decoder on CLIP pre-trained image encoder to predict query masks and corresponding object scores and mask IoU scores. Then, temporal topK-enhanced matching performs query matching across frames using the K mostly matched frames. Finally, weighted open-vocabulary classification first employs mask pooling to generate query visual features from CLIP pre-trained image encoder, and second performs weighted classification using object scores and mask IoU scores. Our CLIP-VIS does not require the annotations of instance categories and identities. The experiments are performed on various video instance segmentation datasets, which demonstrate the effectiveness of our proposed method, especially for novel categories. When using ConvNeXt-B as backbone, our CLIP-VIS achieves the AP and APn scores of 32.2% and 40.2% on the validation set of LV-VIS dataset, which outperforms OV2Seg by 11.1% and 23.9% respectively. We will release the source code and models at https://github.com/zwq456/CLIP-VIS.git.
Human mesh recovery from arbitrary multi-view images involves two characteristics: the arbitrary camera poses and arbitrary number of camera views. Because of the variability, designing a unified framework to tackle this task is challenging. The challenges can be summarized as the dilemma of being able to simultaneously estimate arbitrary camera poses and recover human mesh from arbitrary multi-view images while maintaining flexibility. To solve this dilemma, we propose a divide and conquer framework for Unified Human Mesh Recovery (U-HMR) from arbitrary multi-view images. In particular, U-HMR consists of a decoupled structure and two main components: camera and body decoupling (CBD), camera pose estimation (CPE), and arbitrary view fusion (AVF). As camera poses and human body mesh are independent of each other, CBD splits the estimation of them into two sub-tasks for two individual sub-networks (ie, CPE and AVF) to handle respectively, thus the two sub-tasks are disentangled. In CPE, since each camera pose is unrelated to the others, we adopt a shared MLP to process all views in a parallel way. In AVF, in order to fuse multi-view information and make the fusion operation independent of the number of views, we introduce a transformer decoder with a SMPL parameters query token to extract cross-view features for mesh recovery. To demonstrate the efficacy and flexibility of the proposed framework and effect of each component, we conduct extensive experiments on three public datasets: Human3.6M, MPI-INF-3DHP, and TotalCapture.
Brain tumor image segmentation is a challenging research topic in which deep-learning models have presented the best results. However, the traditional way of training those models from many pre-annotated images leaves several unanswered questions. Hence methodologies, such as Feature Learning from Image Markers (FLIM), have involved an expert in the learning loop to reduce human effort in data annotation and build models sufficiently deep for a given problem. FLIM has been successfully used to create encoders, estimating the filters of all convolutional layers from patches centered at marker voxels. In this work, we present Multi-Step (MS) FLIM - a user-assisted approach to estimating and selecting the most relevant filters from multiple FLIM executions. MS-FLIM is used only for the first convolutional layer, and the results already indicate improvement over FLIM. For evaluation, we build a simple U-shaped encoder-decoder network, named sU-Net, for glioblastoma segmentation using T1Gd and FLAIR MRI scans, varying the encoder's training method, using FLIM, MS-FLIM, and backpropagation algorithm. Also, we compared these sU-Nets with two State-Of-The-Art (SOTA) deep-learning models using two datasets. The results show that the sU-Net based on MS-FLIM outperforms the other training methods and achieves effectiveness within the standard deviations of the SOTA models.
The field of neural rendering has witnessed significant progress with advancements in generative models and differentiable rendering techniques. Though 2D diffusion has achieved success, a unified 3D diffusion pipeline remains unsettled. This paper introduces a novel framework called LN3Diff to address this gap and enable fast, high-quality, and generic conditional 3D generation. Our approach harnesses a 3D-aware architecture and variational autoencoder (VAE) to encode the input image into a structured, compact, and 3D latent space. The latent is decoded by a transformer-based decoder into a high-capacity 3D neural field. Through training a diffusion model on this 3D-aware latent space, our method achieves state-of-the-art performance on ShapeNet for 3D generation and demonstrates superior performance in monocular 3D reconstruction and conditional 3D generation across various datasets. Moreover, it surpasses existing 3D diffusion methods in terms of inference speed, requiring no per-instance optimization. Our proposed LN3Diff presents a significant advancement in 3D generative modeling and holds promise for various applications in 3D vision and graphics tasks.
In this work, we present denoiSplit, a method to tackle a new analysis task, i.e. the challenge of joint semantic image splitting and unsupervised denoising. This dual approach has important applications in fluorescence microscopy, where semantic image splitting has important applications but noise does generally hinder the downstream analysis of image content. Image splitting involves dissecting an image into its distinguishable semantic structures. We show that the current state-of-the-art method for this task struggles in the presence of image noise, inadvertently also distributing the noise across the predicted outputs. The method we present here can deal with image noise by integrating an unsupervised denoising subtask. This integration results in improved semantic image unmixing, even in the presence of notable and realistic levels of imaging noise. A key innovation in denoiSplit is the use of specifically formulated noise models and the suitable adjustment of KL-divergence loss for the high-dimensional hierarchical latent space we are training. We showcase the performance of denoiSplit across multiple tasks on real-world microscopy images. Additionally, we perform qualitative and quantitative evaluations and compare the results to existing benchmarks, demonstrating the effectiveness of using denoiSplit: a single Variational Splitting Encoder-Decoder (VSE) Network using two suitable noise models to jointly perform semantic splitting and denoising.
Recent singing-voice-synthesis (SVS) methods have achieved remarkable audio quality and naturalness, yet they lack the capability to control the style attributes of the synthesized singing explicitly. We propose Prompt-Singer, the first SVS method that enables attribute controlling on singer gender, vocal range and volume with natural language. We adopt a model architecture based on a decoder-only transformer with a multi-scale hierarchy, and design a range-melody decoupled pitch representation that enables text-conditioned vocal range control while keeping melodic accuracy. Furthermore, we explore various experiment settings, including different types of text representations, text encoder fine-tuning, and introducing speech data to alleviate data scarcity, aiming to facilitate further research. Experiments show that our model achieves favorable controlling ability and audio quality. Audio samples are available at http://prompt-singer.github.io .
Streaming text generation has become a common way of increasing the responsiveness of language model powered applications, such as chat assistants. At the same time, extracting semantic information from generated text is a useful tool for applications such as automated fact checking or retrieval augmented generation. Currently, this requires either separate models during inference, which increases computational cost, or destructive fine-tuning of the language model. Instead, we propose an approach called EMBER which enables streaming named entity recognition in decoder-only language models without fine-tuning them and while incurring minimal additional computational cost at inference time. Specifically, our experiments show that EMBER maintains high token generation rates, with only a negligible decrease in speed of around 1% compared to a 43.64% slowdown measured for a baseline. We make our code and data available online, including a toolkit for training, testing, and deploying efficient token classification models optimized for streaming text generation.