Transformers have become a key architecture in speech processing, but our understanding of how they build up representations of acoustic and linguistic structure is limited. In this study, we address this gap by investigating how measures of 'context-mixing' developed for text models can be adapted and applied to models of spoken language. We identify a linguistic phenomenon that is ideal for such a case study: homophony in French (e.g. livre vs livres), where a speech recognition model has to attend to syntactic cues such as determiners and pronouns in order to disambiguate spoken words with identical pronunciations and transcribe them while respecting grammatical agreement. We perform a series of controlled experiments and probing analyses on Transformer-based speech models. Our findings reveal that representations in encoder-only models effectively incorporate these cues to identify the correct transcription, whereas encoders in encoder-decoder models mainly relegate the task of capturing contextual dependencies to decoder modules.
Recent research works have proposed machine learning models for classifying IoT devices connected to a network. However, there is still a practical challenge of not having all devices (and hence their traffic) available during the training of a model. This essentially means, during the operational phase, we need to classify new devices not seen in the training phase. To address this challenge, we propose ZEST -- a ZSL (zero-shot learning) framework based on self-attention for classifying both seen and unseen devices. ZEST consists of i) a self-attention based network feature extractor, termed SANE, for extracting latent space representations of IoT traffic, ii) a generative model that trains a decoder using latent features to generate pseudo data, and iii) a supervised model that is trained on the generated pseudo data for classifying devices. We carry out extensive experiments on real IoT traffic data; our experiments demonstrate i) ZEST achieves significant improvement (in terms of accuracy) over the baselines; ii) SANE is able to better extract meaningful representations than LSTM which has been commonly used for modeling network traffic.
Accurate analysis of microscopy images is hindered by the presence of noise. This noise is usually signal-dependent and often additionally correlated along rows or columns of pixels. Current self- and unsupervised denoisers can address signal-dependent noise, but none can reliably remove noise that is also row- or column-correlated. Here, we present the first fully unsupervised deep learning-based denoiser capable of handling imaging noise that is row-correlated as well as signal-dependent. Our approach uses a Variational Autoencoder (VAE) with a specially designed autoregressive decoder. This decoder is capable of modeling row-correlated and signal-dependent noise but is incapable of independently modeling underlying clean signal. The VAE therefore produces latent variables containing only clean signal information, and these are mapped back into image space using a proposed second decoder network. Our method does not require a pre-trained noise model and can be trained from scratch using unpaired noisy data. We benchmark our approach on microscopy datatsets from a range of imaging modalities and sensor types, each with row- or column-correlated, signal-dependent noise, and show that it outperforms existing self- and unsupervised denoisers.
Medical image segmentation plays a crucial role in advancing healthcare systems for disease diagnosis and treatment planning. The u-shaped architecture, popularly known as U-Net, has proven highly successful for various medical image segmentation tasks. However, U-Net's convolution-based operations inherently limit its ability to model long-range dependencies effectively. To address these limitations, researchers have turned to Transformers, renowned for their global self-attention mechanisms, as alternative architectures. One popular network is our previous TransUNet, which leverages Transformers' self-attention to complement U-Net's localized information with the global context. In this paper, we extend the 2D TransUNet architecture to a 3D network by building upon the state-of-the-art nnU-Net architecture, and fully exploring Transformers' potential in both the encoder and decoder design. We introduce two key components: 1) A Transformer encoder that tokenizes image patches from a convolution neural network (CNN) feature map, enabling the extraction of global contexts, and 2) A Transformer decoder that adaptively refines candidate regions by utilizing cross-attention between candidate proposals and U-Net features. Our investigations reveal that different medical tasks benefit from distinct architectural designs. The Transformer encoder excels in multi-organ segmentation, where the relationship among organs is crucial. On the other hand, the Transformer decoder proves more beneficial for dealing with small and challenging segmented targets such as tumor segmentation. Extensive experiments showcase the significant potential of integrating a Transformer-based encoder and decoder into the u-shaped medical image segmentation architecture. TransUNet outperforms competitors in various medical applications.
Significant progress has been made recently in point cloud segmentation utilizing an encoder-decoder framework, which initially encodes point clouds into low-resolution representations and subsequently decodes high-resolution predictions. Inspired by the success of high-resolution architectures in image dense prediction, which always maintains a high-resolution representation throughout the entire learning process, we consider it also highly important for 3D dense point cloud analysis. Therefore, in this paper, we explore high-resolution architectures for 3D point cloud segmentation. Specifically, we generalize high-resolution architectures using a unified pipeline named PointHR, which includes a knn-based sequence operator for feature extraction and a differential resampling operator to efficiently communicate different resolutions. Additionally, we propose to avoid numerous on-the-fly computations of high-resolution architectures by pre-computing the indices for both sequence and resampling operators. By doing so, we deliver highly competitive high-resolution architectures while capitalizing on the benefits of well-designed point cloud blocks without additional effort. To evaluate these architectures for dense point cloud analysis, we conduct thorough experiments using S3DIS and ScanNetV2 datasets, where the proposed PointHR outperforms recent state-of-the-art methods without any bells and whistles. The source code is available at \url{https://github.com/haibo-qiu/PointHR}.
Over the past years, foundation models have caused a paradigm shift in machine learning due to their unprecedented capabilities for zero-shot and few-shot generalization. However, despite the success of foundation models in modalities such as natural language processing and computer vision, the development of foundation models for time series forecasting has lagged behind. We present Lag-Llama, a general-purpose foundation model for univariate probabilistic time series forecasting based on a decoder-only transformer architecture that uses lags as covariates. Lag-Llama is pretrained on a large corpus of diverse time series data from several domains, and demonstrates strong zero-shot generalization capabilities compared to a wide range of forecasting models on downstream datasets across domains. Moreover, when fine-tuned on relatively small fractions of such previously unseen datasets, Lag-Llama achieves state-of-the-art performance, outperforming prior deep learning approaches, emerging as the best general-purpose model on average. Lag-Llama serves as a strong contender to the current state-of-art in time series forecasting and paves the way for future advancements in foundation models tailored to time series data.
Defect detection is a critical research area in artificial intelligence. Recently, synthetic data-based self-supervised learning has shown great potential on this task. Although many sophisticated synthesizing strategies exist, little research has been done to investigate the robustness of models when faced with different strategies. In this paper, we focus on this issue and find that existing methods are highly sensitive to them. To alleviate this issue, we present a Discrepancy Aware Framework (DAF), which demonstrates robust performance consistently with simple and cheap strategies across different anomaly detection benchmarks. We hypothesize that the high sensitivity to synthetic data of existing self-supervised methods arises from their heavy reliance on the visual appearance of synthetic data during decoding. In contrast, our method leverages an appearance-agnostic cue to guide the decoder in identifying defects, thereby alleviating its reliance on synthetic appearance. To this end, inspired by existing knowledge distillation methods, we employ a teacher-student network, which is trained based on synthesized outliers, to compute the discrepancy map as the cue. Extensive experiments on two challenging datasets prove the robustness of our method. Under the simple synthesis strategies, it outperforms existing methods by a large margin. Furthermore, it also achieves the state-of-the-art localization performance. Code is available at: https://github.com/caiyuxuan1120/DAF.
Pretraining auto-regressive large language models~(LLMs) with retrieval demonstrates better perplexity and factual accuracy by leveraging external databases. However, the size of existing pretrained retrieval-augmented LLM is still limited (e.g., Retro has 7.5B parameters), which limits the effectiveness of instruction tuning and zero-shot generalization. In this work, we introduce Retro 48B, the largest LLM pretrained with retrieval. Specifically, we continue to pretrain a 43B GPT model on additional 100 billion tokens using the Retro augmentation method by retrieving from 1.2 trillion tokens. Notably, the obtained foundation model, Retro 48B, largely outperforms the counterpart GPT 43B trained on 1.2T tokens in terms of perplexity with only 2.58% additional GPU hours, demonstrating the significant scaling potential of the method. After instruction tuning on Retro, InstructRetro demonstrates significant improvement over the instruction tuned GPT on a wide range of zero-shot tasks. Specifically, the average improvement of InstructRetro is 7% over its GPT counterpart across 8 short-form QA and reading comprehension tasks, 10% over GPT across 4 challenging long-form QA tasks, and 16% over GPT across 3 summarization tasks. Surprisingly, we find that one can ablate the encoder from InstructRetro architecture and directly use its decoder backbone, while achieving comparable results. Our results highlight the promising direction to obtain a better GPT decoder through continued pretraining with retrieval before instruction tuning. Our code and checkpoints are publicly available at: https://huggingface.co/nvidia/retro-48b-instruct-4k.
Reconstructing 3D human shapes from 2D images has received increasing attention recently due to its fundamental support for many high-level 3D applications. Compared with natural images, freehand sketches are much more flexible to depict various shapes, providing a high potential and valuable way for 3D human reconstruction. However, such a task is highly challenging. The sparse abstract characteristics of sketches add severe difficulties, such as arbitrariness, inaccuracy, and lacking image details, to the already badly ill-posed problem of 2D-to-3D reconstruction. Although current methods have achieved great success in reconstructing 3D human bodies from a single-view image, they do not work well on freehand sketches. In this paper, we propose a novel sketch-driven multi-faceted decoder network termed SketchBodyNet to address this task. Specifically, the network consists of a backbone and three separate attention decoder branches, where a multi-head self-attention module is exploited in each decoder to obtain enhanced features, followed by a multi-layer perceptron. The multi-faceted decoders aim to predict the camera, shape, and pose parameters, respectively, which are then associated with the SMPL model to reconstruct the corresponding 3D human mesh. In learning, existing 3D meshes are projected via the camera parameters into 2D synthetic sketches with joints, which are combined with the freehand sketches to optimize the model. To verify our method, we collect a large-scale dataset of about 26k freehand sketches and their corresponding 3D meshes containing various poses of human bodies from 14 different angles. Extensive experimental results demonstrate our SketchBodyNet achieves superior performance in reconstructing 3D human meshes from freehand sketches.
Learning representations that generalize to novel compositions of known concepts is crucial for bridging the gap between human and machine perception. One prominent effort is learning object-centric representations, which are widely conjectured to enable compositional generalization. Yet, it remains unclear when this conjecture will be true, as a principled theoretical or empirical understanding of compositional generalization is lacking. In this work, we investigate when compositional generalization is guaranteed for object-centric representations through the lens of identifiability theory. We show that autoencoders that satisfy structural assumptions on the decoder and enforce encoder-decoder consistency will learn object-centric representations that provably generalize compositionally. We validate our theoretical result and highlight the practical relevance of our assumptions through experiments on synthetic image data.
Active learning (AL), which aims to construct an effective training set by iteratively curating the most formative unlabeled data for annotation, has been widely used in low-resource tasks. Most active learning techniques in classification rely on the model's uncertainty or disagreement to choose unlabeled data, suffering from the problem of over-confidence in superficial patterns and a lack of exploration. Inspired by the cognitive processes in which humans deduce and predict through causal information, we take an initial attempt towards integrating rationales into AL and propose a novel Explainable Active Learning framework (XAL) for low-resource text classification, which aims to encourage classifiers to justify their inferences and delve into unlabeled data for which they cannot provide reasonable explanations. Specifically, besides using a pre-trained bi-directional encoder for classification, we employ a pre-trained uni-directional decoder to generate and score the explanation. We further facilitate the alignment of the model with human reasoning preference through a proposed ranking loss. During the selection of unlabeled data, the predicted uncertainty of the encoder and the explanation score of the decoder complement each other as the final metric to acquire informative data. Extensive experiments on six datasets show that XAL achieves consistent improvement over 9 strong baselines. Analysis indicates that the proposed method can generate corresponding explanations for its predictions.
Semantic segmentation tasks naturally require high-resolution information for pixel-wise segmentation and global context information for class prediction. While existing vision transformers demonstrate promising performance, they often utilize high-resolution context modeling, resulting in a computational bottleneck. In this work, we challenge conventional wisdom and introduce the Low-Resolution Self-Attention (LRSA) mechanism to capture global context at a significantly reduced computational cost, i.e., FLOPs. Our approach involves computing self-attention in a fixed low-resolution space regardless of the input image's resolution, with additional 3x3 depth-wise convolutions to capture fine details in the high-resolution space. We demonstrate the effectiveness of our LRSA approach by building the LRFormer, a vision transformer with an encoder-decoder structure. Extensive experiments on the ADE20K, COCO-Stuff, and Cityscapes datasets demonstrate that LRFormer outperforms state-of-the-art models. he code is available at https://github.com/yuhuan-wu/LRFormer.
Talking face generation has a wide range of potential applications in the field of virtual digital humans. However, rendering high-fidelity facial video while ensuring lip synchronization is still a challenge for existing audio-driven talking face generation approaches. To address this issue, we propose HyperLips, a two-stage framework consisting of a hypernetwork for controlling lips and a high-resolution decoder for rendering high-fidelity faces. In the first stage, we construct a base face generation network that uses the hypernetwork to control the encoding latent code of the visual face information over audio. First, FaceEncoder is used to obtain latent code by extracting features from the visual face information taken from the video source containing the face frame.Then, HyperConv, which weighting parameters are updated by HyperNet with the audio features as input, will modify the latent code to synchronize the lip movement with the audio. Finally, FaceDecoder will decode the modified and synchronized latent code into visual face content. In the second stage, we obtain higher quality face videos through a high-resolution decoder. To further improve the quality of face generation, we trained a high-resolution decoder, HRDecoder, using face images and detected sketches generated from the first stage as input.Extensive quantitative and qualitative experiments show that our method outperforms state-of-the-art work with more realistic, high-fidelity, and lip synchronization. Project page: https://semchan.github.io/HyperLips Project/
Text is ubiquitous in our visual world, conveying crucial information, such as in documents, websites, and everyday photographs. In this work, we propose UReader, a first exploration of universal OCR-free visually-situated language understanding based on the Multimodal Large Language Model (MLLM). By leveraging the shallow text recognition ability of the MLLM, we only finetuned 1.2% parameters and the training cost is much lower than previous work following domain-specific pretraining and finetuning paradigms. Concretely, UReader is jointly finetuned on a wide range of Visually-situated Language Understanding tasks via a unified instruction format. To enhance the visual text and semantic understanding, we further apply two auxiliary tasks with the same format, namely text reading and key points generation tasks. We design a shape-adaptive cropping module before the encoder-decoder architecture of MLLM to leverage the frozen low-resolution vision encoder for processing high-resolution images. Without downstream finetuning, our single model achieves state-of-the-art ocr-free performance in 8 out of 10 visually-situated language understanding tasks, across 5 domains: documents, tables, charts, natural images, and webpage screenshots. Codes and instruction-tuning datasets will be released.
We consider low-latency image transmission over a noisy wireless channel when correlated side information is present only at the receiver side (the Wyner-Ziv scenario). In particular, we are interested in developing practical schemes using a data-driven joint source-channel coding (JSCC) approach, which has been previously shown to outperform conventional separation-based approaches in the practical finite blocklength regimes, and to provide graceful degradation with channel quality. We propose a novel neural network architecture that incorporates the decoder-only side information at multiple stages at the receiver side. Our results demonstrate that the proposed method succeeds in integrating the side information, yielding improved performance at all channel conditions in terms of the various quality measures considered here, especially at low channel signal-to-noise ratios (SNRs) and small bandwidth ratios (BRs). We have made the source code of the proposed method public to enable further research, and the reproducibility of the results.
Masked Autoencoders (MAE) have demonstrated promising performance in self-supervised learning for both 2D and 3D computer vision. Nevertheless, existing MAE-based methods still have certain drawbacks. Firstly, the functional decoupling between the encoder and decoder is incomplete, which limits the encoder's representation learning ability. Secondly, downstream tasks solely utilize the encoder, failing to fully leverage the knowledge acquired through the encoder-decoder architecture in the pre-text task. In this paper, we propose Point Regress AutoEncoder (Point-RAE), a new scheme for regressive autoencoders for point cloud self-supervised learning. The proposed method decouples functions between the decoder and the encoder by introducing a mask regressor, which predicts the masked patch representation from the visible patch representation encoded by the encoder and the decoder reconstructs the target from the predicted masked patch representation. By doing so, we minimize the impact of decoder updates on the representation space of the encoder. Moreover, we introduce an alignment constraint to ensure that the representations for masked patches, predicted from the encoded representations of visible patches, are aligned with the masked patch presentations computed from the encoder. To make full use of the knowledge learned in the pre-training stage, we design a new finetune mode for the proposed Point-RAE. Extensive experiments demonstrate that our approach is efficient during pre-training and generalizes well on various downstream tasks. Specifically, our pre-trained models achieve a high accuracy of \textbf{90.28\%} on the ScanObjectNN hardest split and \textbf{94.1\%} accuracy on ModelNet40, surpassing all the other self-supervised learning methods. Our code and pretrained model are public available at: \url{https://github.com/liuyyy111/Point-RAE}.
Ultrasound imaging serves as an effective and non-invasive diagnostic tool commonly employed in clinical examinations. However, the presence of speckle noise in ultrasound images invariably degrades image quality, impeding the performance of subsequent tasks, such as segmentation and classification. Existing methods for speckle noise reduction frequently induce excessive image smoothing or fail to preserve detailed information adequately. In this paper, we propose a complementary global and local knowledge network for ultrasound denoising with fine-grained refinement. Initially, the proposed architecture employs the L-CSwinTransformer as encoder to capture global information, incorporating CNN as decoder to fuse local features. We expand the resolution of the feature at different stages to extract more global information compared to the original CSwinTransformer. Subsequently, we integrate Fine-grained Refinement Block (FRB) within the skip-connection stage to further augment features. We validate our model on two public datasets, HC18 and BUSI. Experimental results demonstrate that our model can achieve competitive performance in both quantitative metrics and visual performance. Our code will be available at https://github.com/AAlkaid/USDenoising.
Recent advancements in subject-driven image generation have made significant strides. However, current methods still fall short in diverse application scenarios, as they require test-time tuning and cannot accept interleaved multi-image and text input. These limitations keep them far from the ultimate goal of "image as a foreign language in image generation." This paper presents Kosmos-G, a model that leverages the advanced multimodal perception capabilities of Multimodal Large Language Models (MLLMs) to tackle the aforementioned challenge. Our approach aligns the output space of MLLM with CLIP using the textual modality as an anchor and performs compositional instruction tuning on curated data. Kosmos-G demonstrates an impressive capability of zero-shot subject-driven generation with interleaved multi-image and text input. Notably, the score distillation instruction tuning requires no modifications to the image decoder. This allows for a seamless substitution of CLIP and effortless integration with a myriad of U-Net techniques ranging from fine-grained controls to personalized image decoder variants. We posit Kosmos-G as an initial attempt towards the goal of "image as a foreign language in image generation." The code can be found at https://aka.ms/Kosmos-G
This paper presents a classification framework based on learnable data augmentation to tackle the One-Shot Unsupervised Domain Adaptation (OS-UDA) problem. OS-UDA is the most challenging setting in Domain Adaptation, as only one single unlabeled target sample is assumed to be available for model adaptation. Driven by such single sample, our method LearnAug-UDA learns how to augment source data, making it perceptually similar to the target. As a result, a classifier trained on such augmented data will generalize well for the target domain. To achieve this, we designed an encoder-decoder architecture that exploits a perceptual loss and style transfer strategies to augment the source data. Our method achieves state-of-the-art performance on two well-known Domain Adaptation benchmarks, DomainNet and VisDA. The project code is available at https://github.com/IIT-PAVIS/LearnAug-UDA
Current Large Multimodal Models (LMMs) achieve remarkable progress, yet there remains significant uncertainty regarding their ability to accurately apprehend visual details, that is, in performing detailed captioning. To address this, we introduce $\textit{CCEval}$, a GPT-4 assisted evaluation method for detailed captioning. Interestingly, while LMMs demonstrate minimal object existence hallucination in existing VQA benchmarks, our proposed evaluation reveals continued susceptibility to such hallucinations. In this paper, we make the first attempt to investigate such hallucination from different aspects, including image resolution, the language decoder size, and instruction data amount, quality, granularity. Our findings underscore the unwarranted inference when the language description includes details at a finer object granularity than what the vision module can ground or verify, thus inducing hallucination. To control such hallucinations, we further attribute the reliability of captioning to contextual knowledge (involving only contextually grounded objects) and parametric knowledge (containing inferred objects by the model). Thus, we introduce $\textit{HallE-Control}$, a controllable LMM in terms of $\textbf{Hall}$ucination in object $\textbf{E}$xistence. HallE-Control can condition the captioning to shift between (i) exclusively depicting contextual knowledge for grounded objects and (ii) blending it with parametric knowledge to imagine inferred objects. Our method reduces hallucination by 44% compared to LLaVA$_{7B}$ and maintains the object coverage.