We propose to utilize an instruction-tuned large language model (LLM) for guiding the text generation process in automatic speech recognition (ASR). Modern large language models (LLMs) are adept at performing various text generation tasks through zero-shot learning, prompted with instructions designed for specific objectives. This paper explores the potential of LLMs to derive linguistic information that can facilitate text generation in end-to-end ASR models. Specifically, we instruct an LLM to correct grammatical errors in an ASR hypothesis and use the LLM-derived representations to refine the output further. The proposed model is built on the joint CTC and attention architecture, with the LLM serving as a front-end feature extractor for the decoder. The ASR hypothesis, subject to correction, is obtained from the encoder via CTC decoding and fed into the LLM along with a specific instruction. The decoder subsequently takes as input the LLM output to perform token predictions, combining acoustic information from the encoder and the powerful linguistic information provided by the LLM. Experimental results show that the proposed LLM-guided model achieves a relative gain of approximately 13\% in word error rates across major benchmarks.
Semantic segmentation is an essential technology for self-driving cars to comprehend their surroundings. Currently, real-time semantic segmentation networks commonly employ either encoder-decoder architecture or two-pathway architecture. Generally speaking, encoder-decoder models tend to be quicker,whereas two-pathway models exhibit higher accuracy. To leverage both strengths, we present the Spatial-Assistant Encoder-Decoder Network (SANet) to fuse the two architectures. In the overall architecture, we uphold the encoder-decoder design while maintaining the feature maps in the middle section of the encoder and utilizing atrous convolution branches for same-resolution feature extraction. Toward the end of the encoder, we integrate the asymmetric pooling pyramid pooling module (APPPM) to optimize the semantic extraction of the feature maps. This module incorporates asymmetric pooling layers that extract features at multiple resolutions. In the decoder, we present a hybrid attention module, SAD, that integrates horizontal and vertical attention to facilitate the combination of various branches. To ascertain the effectiveness of our approach, our SANet model achieved competitive results on the real-time CamVid and cityscape datasets. By employing a single 2080Ti GPU, SANet achieved a 78.4 % mIOU at 65.1 FPS on the Cityscape test dataset and 78.8 % mIOU at 147 FPS on the CamVid test dataset. The training code and model for SANet are available at https://github.com/CuZaoo/SANet-main
We introduce the elEmBERT model for chemical classification tasks. It is based on deep learning techniques, such as a multilayer encoder architecture. We demonstrate the opportunities offered by our approach on sets of organic, inorganic and crystalline compounds. In particular, we developed and tested the model using the Matbench and Moleculenet benchmarks, which include crystal properties and drug design-related benchmarks. We also conduct an analysis of vector representations of chemical compounds, shedding light on the underlying patterns in structural data. Our model exhibits exceptional predictive capabilities and proves universally applicable to molecular and material datasets. For instance, on the Tox21 dataset, we achieved an average precision of 96%, surpassing the previously best result by 10%.
Digital video splicing has become easy and ubiquitous. Malicious users copy some regions of a video and paste them to another video for creating realistic forgeries. It is significant to blindly detect such forgery regions in videos. In this paper, a spatio-temporal co-attention fusion network (SCFNet) is proposed for video splicing localization. Specifically, a three-stream network is used as an encoder to capture manipulation traces across multiple frames. The deep interaction and fusion of spatio-temporal forensic features are achieved by the novel parallel and cross co-attention fusion modules. A lightweight multilayer perceptron (MLP) decoder is adopted to yield a pixel-level tampering localization map. A new large-scale video splicing dataset is created for training the SCFNet. Extensive tests on benchmark datasets show that the localization and generalization performances of our SCFNet outperform the state-of-the-art. Code and datasets will be available at https://github.com/multimediaFor/SCFNet.
With the frequent use of self-supervised monocular depth estimation in robotics and autonomous driving, the model's efficiency is becoming increasingly important. Most current approaches apply much larger and more complex networks to improve the precision of depth estimation. Some researchers incorporated Transformer into self-supervised monocular depth estimation to achieve better performance. However, this method leads to high parameters and high computation. We present a fully convolutional depth estimation network using contextual feature fusion. Compared to UNet++ and HRNet, we use high-resolution and low-resolution features to reserve information on small targets and fast-moving objects instead of long-range fusion. We further promote depth estimation results employing lightweight channel attention based on convolution in the decoder stage. Our method reduces the parameters without sacrificing accuracy. Experiments on the KITTI benchmark show that our method can get better results than many large models, such as Monodepth2, with only 30 parameters. The source code is available at https://github.com/boyagesmile/DNA-Depth.
Recent end-to-end automatic speech recognition (ASR) models have become increasingly larger, making them particularly challenging to be deployed on resource-constrained devices. Model quantisation is an effective solution that sometimes causes the word error rate (WER) to increase. In this paper, a novel strategy of personalisation for a quantised model (PQM) is proposed, which combines speaker adaptive training (SAT) with model quantisation to improve the performance of heavily compressed models. Specifically, PQM uses a 4-bit NormalFloat Quantisation (NF4) approach for model quantisation and low-rank adaptation (LoRA) for SAT. Experiments have been performed on the LibriSpeech and the TED-LIUM 3 corpora. Remarkably, with a 7x reduction in model size and 1% additional speaker-specific parameters, 15.1% and 23.3% relative WER reductions were achieved on quantised Whisper and Conformer-based attention-based encoder-decoder ASR models respectively, comparing to the original full precision models.
We present RECAP (REtrieval-Augmented Audio CAPtioning), a novel and effective audio captioning system that generates captions conditioned on an input audio and other captions similar to the audio retrieved from a datastore. Additionally, our proposed method can transfer to any domain without the need for any additional fine-tuning. To generate a caption for an audio sample, we leverage an audio-text model CLAP to retrieve captions similar to it from a replaceable datastore, which are then used to construct a prompt. Next, we feed this prompt to a GPT-2 decoder and introduce cross-attention layers between the CLAP encoder and GPT-2 to condition the audio for caption generation. Experiments on two benchmark datasets, Clotho and AudioCaps, show that RECAP achieves competitive performance in in-domain settings and significant improvements in out-of-domain settings. Additionally, due to its capability to exploit a large text-captions-only datastore in a training-free fashion, RECAP shows unique capabilities of captioning novel audio events never seen during training and compositional audios with multiple events. To promote research in this space, we also release 150,000+ new weakly labeled captions for AudioSet, AudioCaps, and Clotho.
In this work, we propose an error correction framework, named DiaCorrect, to refine the output of a diarization system in a simple yet effective way. This method is inspired by error correction techniques in automatic speech recognition. Our model consists of two parallel convolutional encoders and a transform-based decoder. By exploiting the interactions between the input recording and the initial system's outputs, DiaCorrect can automatically correct the initial speaker activities to minimize the diarization errors. Experiments on 2-speaker telephony data show that the proposed DiaCorrect can effectively improve the initial model's results. Our source code is publicly available at https://github.com/BUTSpeechFIT/diacorrect.
Modern approaches have proved the huge potential of addressing semantic segmentation as a mask classification task which is widely used in instance-level segmentation. This paradigm trains models by assigning part of object queries to ground truths via conventional one-to-one matching. However, we observe that the popular video semantic segmentation (VSS) dataset has limited categories per video, meaning less than 10% of queries could be matched to receive meaningful gradient updates during VSS training. This inefficiency limits the full expressive potential of all queries.Thus, we present a novel solution THE-Mask for VSS, which introduces temporal-aware hierarchical object queries for the first time. Specifically, we propose to use a simple two-round matching mechanism to involve more queries matched with minimal cost during training while without any extra cost during inference. To support our more-to-one assignment, in terms of the matching results, we further design a hierarchical loss to train queries with their corresponding hierarchy of primary or secondary. Moreover, to effectively capture temporal information across frames, we propose a temporal aggregation decoder that fits seamlessly into the mask-classification paradigm for VSS. Utilizing temporal-sensitive multi-level queries, our method achieves state-of-the-art performance on the latest challenging VSS benchmark VSPW without bells and whistles.
Never having seen an object and heard its sound simultaneously, can the model still accurately localize its visual position from the input audio? In this work, we concentrate on the Audio-Visual Localization and Segmentation tasks but under the demanding zero-shot and few-shot scenarios. To achieve this goal, different from existing approaches that mostly employ the encoder-fusion-decoder paradigm to decode localization information from the fused audio-visual feature, we introduce the encoder-prompt-decoder paradigm, aiming to better fit the data scarcity and varying data distribution dilemmas with the help of abundant knowledge from pre-trained models. Specifically, we first propose to construct Semantic-aware Audio Prompt (SAP) to help the visual foundation model focus on sounding objects, meanwhile, the semantic gap between the visual and audio modalities is also encouraged to shrink. Then, we develop a Correlation Adapter (ColA) to keep minimal training efforts as well as maintain adequate knowledge of the visual foundation model. By equipping with these means, extensive experiments demonstrate that this new paradigm outperforms other fusion-based methods in both the unseen class and cross-dataset settings. We hope that our work can further promote the generalization study of Audio-Visual Localization and Segmentation in practical application scenarios.
Real-time transportation surveillance is an essential part of the intelligent transportation system (ITS). However, images captured under low-light conditions often suffer the poor visibility with types of degradation, such as noise interference and vague edge features, etc. With the development of imaging devices, the quality of the visual surveillance data is continually increasing, like 2K and 4K, which has more strict requirements on the efficiency of image processing. To satisfy the requirements on both enhancement quality and computational speed, this paper proposes a double domain guided real-time low-light image enhancement network (DDNet) for ultra-high-definition (UHD) transportation surveillance. Specifically, we design an encoder-decoder structure as the main architecture of the learning network. In particular, the enhancement processing is divided into two subtasks (i.e., color enhancement and gradient enhancement) via the proposed coarse enhancement module (CEM) and LoG-based gradient enhancement module (GEM), which are embedded in the encoder-decoder structure. It enables the network to enhance the color and edge features simultaneously. Through the decomposition and reconstruction on both color and gradient domains, our DDNet can restore the detailed feature information concealed by the darkness with better visual quality and efficiency. The evaluation experiments on standard and transportation-related datasets demonstrate that our DDNet provides superior enhancement quality and efficiency compared with the state-of-the-art methods. Besides, the object detection and scene segmentation experiments indicate the practical benefits for higher-level image analysis under low-light environments in ITS.
This paper works on non-autoregressive automatic speech recognition. A unimodal aggregation (UMA) is proposed to segment and integrate the feature frames that belong to the same text token, and thus to learn better feature representations for text tokens. The frame-wise features and weights are both derived from an encoder. Then, the feature frames with unimodal weights are integrated and further processed by a decoder. Connectionist temporal classification (CTC) loss is applied for training. Compared to the regular CTC, the proposed method learns better feature representations and shortens the sequence length, resulting in lower recognition error and computational complexity. Experiments on three Mandarin datasets show that UMA demonstrates superior or comparable performance to other advanced non-autoregressive methods, such as self-conditioned CTC. Moreover, by integrating self-conditioned CTC into the proposed framework, the performance can be further noticeably improved.
Automated Audio Captioning (AAC) is the task of generating natural language descriptions given an audio stream. A typical AAC system requires manually curated training data of audio segments and corresponding text caption annotations. The creation of these audio-caption pairs is costly, resulting in general data scarcity for the task. In this work, we address this major limitation and propose an approach to train AAC systems using only text. Our approach leverages the multimodal space of contrastively trained audio-text models, such as CLAP. During training, a decoder generates captions conditioned on the pretrained CLAP text encoder. During inference, the text encoder is replaced with the pretrained CLAP audio encoder. To bridge the modality gap between text and audio embeddings, we propose the use of noise injection or a learnable adapter, during training. We find that the proposed text-only framework performs competitively with state-of-the-art models trained with paired audio, showing that efficient text-to-audio transfer is possible. Finally, we showcase both stylized audio captioning and caption enrichment while training without audio or human-created text captions.
The task of blind source separation (BSS) involves separating sources from a mixture without prior knowledge of the sources or the mixing system. Single-channel mixtures and non-linear mixtures are a particularly challenging problem in BSS. In this paper, we propose a novel method for addressing BSS with single-channel non-linear mixtures by leveraging the natural feature subspace specialization ability of multi-encoder autoencoders. During the training phase, our method unmixes the input into the separate encoding spaces of the multi-encoder network and then remixes these representations within the decoder for a reconstruction of the input. Then to perform source inference, we introduce a novel encoding masking technique whereby masking out all but one of the encodings enables the decoder to estimate a source signal. To this end, we also introduce a sparse mixing loss that encourages sparse remixing of source encodings throughout the decoder and a so-called zero reconstruction loss on the decoder for coherent source estimations. To analyze and evaluate our method, we conduct experiments on a toy dataset, designed to demonstrate this property of feature subspace specialization, and with real-world biosignal recordings from a polysomnography sleep study for extracting respiration from electrocardiogram and photoplethysmography signals.
To enhance the reasoning capabilities of off-the-shelf Large Language Models (LLMs), we introduce a simple, yet general and effective prompting method, Re2, i.e., \textbf{Re}-\textbf{Re}ading the question as input. Unlike most thought-eliciting prompting methods, such as Chain-of-Thought (CoT), which aim to elicit the reasoning process in the output, Re2 shifts the focus to the input by processing questions twice, thereby enhancing the understanding process. Consequently, Re2 demonstrates strong generality and compatibility with most thought-eliciting prompting methods, including CoT. Crucially, Re2 facilitates a "bidirectional" encoding in unidirectional decoder-only LLMs because the first pass could provide global information for the second pass. We begin with a preliminary empirical study as the foundation of Re2, illustrating its potential to enable "bidirectional" attention mechanisms. We then evaluate Re2 on extensive reasoning benchmarks across 14 datasets, spanning 112 experiments, to validate its effectiveness and generality. Our findings indicate that, with the exception of a few scenarios on vanilla ChatGPT, Re2 consistently enhances the reasoning performance of LLMs through a simple re-reading strategy. Further analyses reveal Re2's adaptability, showing how it can be effectively integrated with different LLMs, thought-eliciting prompting, and ensemble strategies. Our code is available at \url{https://github.com/Tebmer/Rereading-LLM-Reasoning/}
Attribute reconstruction is used to predict node or edge features in the pre-training of graph neural networks. Given a large number of molecules, they learn to capture structural knowledge, which is transferable for various downstream property prediction tasks and vital in chemistry, biomedicine, and material science. Previous strategies that randomly select nodes to do attribute masking leverage the information of local neighbors However, the over-reliance of these neighbors inhibits the model's ability to learn from higher-level substructures. For example, the model would learn little from predicting three carbon atoms in a benzene ring based on the other three but could learn more from the inter-connections between the functional groups, or called chemical motifs. In this work, we propose and investigate motif-aware attribute masking strategies to capture inter-motif structures by leveraging the information of atoms in neighboring motifs. Once each graph is decomposed into disjoint motifs, the features for every node within a sample motif are masked. The graph decoder then predicts the masked features of each node within the motif for reconstruction. We evaluate our approach on eight molecular property prediction datasets and demonstrate its advantages.
Depth estimation plays an important role in the robotic perception system. Self-supervised monocular paradigm has gained significant attention since it can free training from the reliance on depth annotations. Despite recent advancements, existing self-supervised methods still underutilize the available training data, limiting their generalization ability. In this paper, we take two data augmentation techniques, namely Resizing-Cropping and Splitting-Permuting, to fully exploit the potential of training datasets. Specifically, the original image and the generated two augmented images are fed into the training pipeline simultaneously and we leverage them to conduct self-distillation. Additionally, we introduce the detail-enhanced DepthNet with an extra full-scale branch in the encoder and a grid decoder to enhance the restoration of fine details in depth maps. Experimental results demonstrate our method can achieve state-of-the-art performance on the KITTI benchmark, with both raw ground truth and improved ground truth. Moreover, our models also show superior generalization performance when transferring to Make3D and NYUv2 datasets. Our codes are available at https://github.com/Sauf4896/BDEdepth.
Since rain streaks show a variety of shapes and directions, learning the degradation representation is extremely challenging for single image deraining. Existing methods are mainly targeted at designing complicated modules to implicitly learn latent degradation representation from coupled rainy images. This way, it is hard to decouple the content-independent degradation representation due to the lack of explicit constraint, resulting in over- or under-enhancement problems. To tackle this issue, we propose a novel Latent Degradation Representation Constraint Network (LDRCNet) that consists of Direction-Aware Encoder (DAEncoder), UNet Deraining Network, and Multi-Scale Interaction Block (MSIBlock). Specifically, the DAEncoder is proposed to adaptively extract latent degradation representation by using the deformable convolutions to exploit the direction consistency of rain streaks. Next, a constraint loss is introduced to explicitly constraint the degradation representation learning during training. Last, we propose an MSIBlock to fuse with the learned degradation representation and decoder features of the deraining network for adaptive information interaction, which enables the deraining network to remove various complicated rainy patterns and reconstruct image details. Experimental results on synthetic and real datasets demonstrate that our method achieves new state-of-the-art performance.
Detecting breast lesion in videos is crucial for computer-aided diagnosis. Existing video-based breast lesion detection approaches typically perform temporal feature aggregation of deep backbone features based on the self-attention operation. We argue that such a strategy struggles to effectively perform deep feature aggregation and ignores the useful local information. To tackle these issues, we propose a spatial-temporal deformable attention based framework, named STNet. Our STNet introduces a spatial-temporal deformable attention module to perform local spatial-temporal feature fusion. The spatial-temporal deformable attention module enables deep feature aggregation in each stage of both encoder and decoder. To further accelerate the detection speed, we introduce an encoder feature shuffle strategy for multi-frame prediction during inference. In our encoder feature shuffle strategy, we share the backbone and encoder features, and shuffle encoder features for decoder to generate the predictions of multiple frames. The experiments on the public breast lesion ultrasound video dataset show that our STNet obtains a state-of-the-art detection performance, while operating twice as fast inference speed. The code and model are available at https://github.com/AlfredQin/STNet.
Accurate medical image segmentation especially for echocardiographic images with unmissable noise requires elaborate network design. Compared with manual design, Neural Architecture Search (NAS) realizes better segmentation results due to larger search space and automatic optimization, but most of the existing methods are weak in layer-wise feature aggregation and adopt a ``strong encoder, weak decoder" structure, insufficient to handle global relationships and local details. To resolve these issues, we propose a novel semi-supervised hybrid NAS network for accurate medical image segmentation termed SSHNN. In SSHNN, we creatively use convolution operation in layer-wise feature fusion instead of normalized scalars to avoid losing details, making NAS a stronger encoder. Moreover, Transformers are introduced for the compensation of global context and U-shaped decoder is designed to efficiently connect global context with local features. Specifically, we implement a semi-supervised algorithm Mean-Teacher to overcome the limited volume problem of labeled medical image dataset. Extensive experiments on CAMUS echocardiography dataset demonstrate that SSHNN outperforms state-of-the-art approaches and realizes accurate segmentation. Code will be made publicly available.