Large language models (LLMs) have demonstrated remarkable abilities in representation learning for program synthesis and understanding tasks. The quality of the learned representations appears to be dictated by the neural scaling laws as a function of the number of model parameters and observations, while imposing upper bounds on the model performance by the amount of available data and compute, which is costly. In this study, we attempt to render the training of LLMs for program synthesis more efficient by unifying four key components: (1) model architectures, (2) learning methods, (3) infill sampling, and, (4) data distributions. Specifically, for the model architecture, we attempt to unify encoder and decoder-based models into a single prefix-LM. For learning methods, (i) causal language modeling, (ii) span corruption, (iii) infilling are unified into a simple learning algorithm. For infill sampling, we explore the claim of a "free lunch" hypothesis. For data distributions, the effect of a mixture distribution and multi-epoch training of programming and natural languages on model performance is explored. We conduct a comprehensive series of empirical experiments on 1B LLMs, for which failures and successes of this exploration are distilled into five lessons. We will provide a final recipe for training and release CodeGen2 models in size 1B, 3.7B, 7B, and, 16B parameters, along with the training framework as open-source: https://github.com/salesforce/CodeGen.
Segment Anything Model (SAM) has recently shown its powerful effectiveness in visual segmentation tasks. However, there is less exploration concerning how SAM works on audio-visual tasks, such as visual sound localization and segmentation. In this work, we propose a simple yet effective audio-visual localization and segmentation framework based on the Segment Anything Model, namely AV-SAM, that can generate sounding object masks corresponding to the audio. Specifically, our AV-SAM simply leverages pixel-wise audio-visual fusion across audio features and visual features from the pre-trained image encoder in SAM to aggregate cross-modal representations. Then, the aggregated cross-modal features are fed into the prompt encoder and mask decoder to generate the final audio-visual segmentation masks. We conduct extensive experiments on Flickr-SoundNet and AVSBench datasets. The results demonstrate that the proposed AV-SAM can achieve competitive performance on sounding object localization and segmentation.
We propose to Transform Scene Graphs (TSG) into more descriptive captions. In TSG, we apply multi-head attention (MHA) to design the Graph Neural Network (GNN) for embedding scene graphs. After embedding, different graph embeddings contain diverse specific knowledge for generating the words with different part-of-speech, e.g., object/attribute embedding is good for generating nouns/adjectives. Motivated by this, we design a Mixture-of-Expert (MOE)-based decoder, where each expert is built on MHA, for discriminating the graph embeddings to generate different kinds of words. Since both the encoder and decoder are built based on the MHA, as a result, we construct a homogeneous encoder-decoder unlike the previous heterogeneous ones which usually apply Fully-Connected-based GNN and LSTM-based decoder. The homogeneous architecture enables us to unify the training configuration of the whole model instead of specifying different training strategies for diverse sub-networks as in the heterogeneous pipeline, which releases the training difficulty. Extensive experiments on the MS-COCO captioning benchmark validate the effectiveness of our TSG. The code is in: https://github.com/GaryJiajia/TSG.
Flight Trajectory Prediction (FTP) is an essential task in Air Traffic Control (ATC), which can assist air traffic controllers in managing airspace more safely and efficiently. Existing approaches generally perform multi-horizon FTP tasks in an autoregressive manner, thereby suffering from error accumulation and low-efficiency problems. In this paper, a novel framework, called FlightBERT++, is proposed to i) forecast multi-horizon flight trajectories directly in a non-autoregressive way, and ii) improve the limitation of the binary encoding (BE) representation in the FlightBERT framework. Specifically, the proposed framework is implemented by a generalized encoder-decoder architecture, in which the encoder learns the temporal-spatial patterns from historical observations and the decoder predicts the flight status for the future horizons. Compared to conventional architecture, an innovative horizon-aware contexts generator is dedicatedly designed to consider the prior horizon information, which further enables non-autoregressive multi-horizon prediction. Additionally, the Gray code representation and the differential prediction paradigm are designed to cope with the high-bit misclassifications of the BE representation, which significantly reduces the outliers in the predictions. Moreover, a differential prompted decoder is proposed to enhance the capability of the differential predictions by leveraging the stationarity of the differential sequence. Extensive experiments are conducted to validate the proposed framework on a real-world flight trajectory dataset. The experimental results demonstrated that the proposed framework outperformed the competitive baselines in both FTP performance and computational efficiency.
Low-Power Wide-Area Networks (LPWANs) are an emerging Internet-of-Things (IoT) paradigm marked by low-power and long-distance communication. Among them, LoRa is widely deployed for its unique characteristics and open-source technology. By adopting the Chirp Spread Spectrum (CSS) modulation, LoRa enables low signal-to-noise ratio (SNR) communication. The standard LoRa demodulation method accumulates the chirp power of the whole chirp into an energy peak in the frequency domain. In this way, it can support communication even when SNR is lower than -15 dB. Beyond that, we proposed NELoRa, a neural-enhanced decoder that exploits multi-dimensional information to achieve significant SNR gain. This paper presents the dataset used to train/test NELoRa, which includes 27,329 LoRa symbols with spreading factors from 7 to 10, for further improvement of neural-enhanced LoRa demodulation. The dataset shows that NELoRa can achieve 1.84-2.35 dB SNR gain over the standard LoRa decoder. The dataset and codes can be found at https://github.com/daibiaoxuwu/NeLoRa_Dataset.
Recent research demonstrates that external knowledge injection can advance pre-trained language models (PLMs) in a variety of downstream NLP tasks. However, existing knowledge injection methods are either applicable to structured knowledge or unstructured knowledge, lacking a unified usage. In this paper, we propose a UNified knowledge inTERface, UNTER, to provide a unified perspective to exploit both structured knowledge and unstructured knowledge. In UNTER, we adopt the decoder as a unified knowledge interface, aligning span representations obtained from the encoder with their corresponding knowledge. This approach enables the encoder to uniformly invoke span-related knowledge from its parameters for downstream applications. Experimental results show that, with both forms of knowledge injected, UNTER gains continuous improvements on a series of knowledge-driven NLP tasks, including entity typing, named entity recognition and relation extraction, especially in low-resource scenarios.
Cross-lingual and cross-domain knowledge alignment without sufficient external resources is a fundamental and crucial task for fusing irregular data. As the element-wise fusion process aiming to discover equivalent objects from different knowledge graphs (KGs), entity alignment (EA) has been attracting great interest from industry and academic research recent years. Most of existing EA methods usually explore the correlation between entities and relations through neighbor nodes, structural information and external resources. However, the complex intrinsic interactions among triple elements and role information are rarely modeled in these methods, which may lead to the inadequate illustration for triple. In addition, external resources are usually unavailable in some scenarios especially cross-lingual and cross-domain applications, which reflects the little scalability of these methods. To tackle the above insufficiency, a novel universal EA framework (OTIEA) based on ontology pair and role enhancement mechanism via triple-aware attention is proposed in this paper without introducing external resources. Specifically, an ontology-enhanced triple encoder is designed via mining intrinsic correlations and ontology pair information instead of independent elements. In addition, the EA-oriented representations can be obtained in triple-aware entity decoder by fusing role diversity. Finally, a bidirectional iterative alignment strategy is deployed to expand seed entity pairs. The experimental results on three real-world datasets show that our framework achieves a competitive performance compared with baselines.
Extensive work has demonstrated the effectiveness of Vision Transformers. The plain Vision Transformer tends to obtain multi-scale features by selecting fixed layers, or the last layer of features aiming to achieve higher performance in dense prediction tasks. However, this selection is often based on manual operation. And different samples often exhibit different features at different layers (e.g., edge, structure, texture, detail, etc.). This requires us to seek a dynamic adaptive fusion method to filter different layer features. In this paper, unlike previous encoder and decoder work, we design a neck network for adaptive fusion and feature selection, called ViTController. We validate the effectiveness of our method on different datasets and models and surpass previous state-of-the-art methods. Finally, our method can also be used as a plug-in module and inserted into different networks.
Crude oil is an integral component of the modern world economy. With the growing demand for crude oil due to its widespread applications, accidental oil spills are unavoidable. Even though oil spills are in and themselves difficult to clean up, the first and foremost challenge is to detect spills. In this research, the authors test the feasibility of deep encoder-decoder models that can be trained effectively to detect oil spills. The work compares the results from several segmentation models on high dimensional satellite Synthetic Aperture Radar (SAR) image data. Multiple combinations of models are used in running the experiments. The best-performing model is the one with the ResNet-50 encoder and DeepLabV3+ decoder. It achieves a mean Intersection over Union (IoU) of 64.868% and a class IoU of 61.549% for the "oil spill" class when compared with the current benchmark model, which achieved a mean IoU of 65.05% and a class IoU of 53.38% for the "oil spill" class.
In this paper, we propose a novel hybrid representation and end-to-end trainable network architecture to model fully editable and customizable neural avatars. At the core of our work lies a representation that combines the modeling power of neural fields with the ease of use and inherent 3D consistency of skinned meshes. To this end, we construct a trainable feature codebook to store local geometry and texture features on the vertices of a deformable body model, thus exploiting its consistent topology under articulation. This representation is then employed in a generative auto-decoder architecture that admits fitting to unseen scans and sampling of realistic avatars with varied appearances and geometries. Furthermore, our representation allows local editing by swapping local features between 3D assets. To verify our method for avatar creation and editing, we contribute a new high-quality dataset, dubbed CustomHumans, for training and evaluation. Our experiments quantitatively and qualitatively show that our method generates diverse detailed avatars and achieves better model fitting performance compared to state-of-the-art methods. Our code and dataset are available at https://custom-humans.github.io/.
Despite radar's popularity in the automotive industry, for fusion-based 3D object detection, most existing works focus on LiDAR and camera fusion. In this paper, we propose TransCAR, a Transformer-based Camera-And-Radar fusion solution for 3D object detection. Our TransCAR consists of two modules. The first module learns 2D features from surround-view camera images and then uses a sparse set of 3D object queries to index into these 2D features. The vision-updated queries then interact with each other via transformer self-attention layer. The second module learns radar features from multiple radar scans and then applies transformer decoder to learn the interactions between radar features and vision-updated queries. The cross-attention layer within the transformer decoder can adaptively learn the soft-association between the radar features and vision-updated queries instead of hard-association based on sensor calibration only. Finally, our model estimates a bounding box per query using set-to-set Hungarian loss, which enables the method to avoid non-maximum suppression. TransCAR improves the velocity estimation using the radar scans without temporal information. The superior experimental results of our TransCAR on the challenging nuScenes datasets illustrate that our TransCAR outperforms state-of-the-art Camera-Radar fusion-based 3D object detection approaches.
Unsupervised learning of object-centric representations in dynamic visual scenes is challenging. Unlike most previous approaches that learn to decompose 2D images, we present DynaVol, a 3D scene generative model that unifies geometric structures and object-centric learning in a differentiable volume rendering framework. The key idea is to perform object-centric voxelization to capture the 3D nature of the scene, which infers the probability distribution over objects at individual spatial locations. These voxel features evolve over time through a canonical-space deformation function, forming the basis for global representation learning via slot attention. The voxel features and global features are complementary and are both leveraged by a compositional NeRF decoder for volume rendering. DynaVol remarkably outperforms existing approaches for unsupervised dynamic scene decomposition. Once trained, the explicitly meaningful voxel features enable additional capabilities that 2D scene decomposition methods cannot achieve: it is possible to freely edit the geometric shapes or manipulate the motion trajectories of the objects.
Medical image segmentation is a fundamental task in the community of medical image analysis. In this paper, a novel network architecture, referred to as Convolution, Transformer, and Operator (CTO), is proposed. CTO employs a combination of Convolutional Neural Networks (CNNs), Vision Transformer (ViT), and an explicit boundary detection operator to achieve high recognition accuracy while maintaining an optimal balance between accuracy and efficiency. The proposed CTO follows the standard encoder-decoder segmentation paradigm, where the encoder network incorporates a popular CNN backbone for capturing local semantic information, and a lightweight ViT assistant for integrating long-range dependencies. To enhance the learning capacity on boundary, a boundary-guided decoder network is proposed that uses a boundary mask obtained from a dedicated boundary detection operator as explicit supervision to guide the decoding learning process. The performance of the proposed method is evaluated on six challenging medical image segmentation datasets, demonstrating that CTO achieves state-of-the-art accuracy with a competitive model complexity.
The lightweight MLP-based decoder has become increasingly promising for semantic segmentation. However, the channel-wise MLP cannot expand the receptive fields, lacking the context modeling capacity, which is critical to semantic segmentation. In this paper, we propose a parametric-free patch rotate operation to reorganize the pixels spatially. It first divides the feature map into multiple groups and then rotates the patches within each group. Based on the proposed patch rotate operation, we design a novel segmentation network, named PRSeg, which includes an off-the-shelf backbone and a lightweight Patch Rotate MLP decoder containing multiple Dynamic Patch Rotate Blocks (DPR-Blocks). In each DPR-Block, the fully connected layer is performed following a Patch Rotate Module (PRM) to exchange spatial information between pixels. Specifically, in PRM, the feature map is first split into the reserved part and rotated part along the channel dimension according to the predicted probability of the Dynamic Channel Selection Module (DCSM), and our proposed patch rotate operation is only performed on the rotated part. Extensive experiments on ADE20K, Cityscapes and COCO-Stuff 10K datasets prove the effectiveness of our approach. We expect that our PRSeg can promote the development of MLP-based decoder in semantic segmentation.
With the increasing demand for video understanding, video moment and highlight detection (MHD) has emerged as a critical research topic. MHD aims to localize all moments and predict clip-wise saliency scores simultaneously. Despite progress made by existing DETR-based methods, we observe that these methods coarsely fuse features from different modalities, which weakens the temporal intra-modal context and results in insufficient cross-modal interaction. To address this issue, we propose MH-DETR (Moment and Highlight Detection Transformer) tailored for MHD. Specifically, we introduce a simple yet efficient pooling operator within the uni-modal encoder to capture global intra-modal context. Moreover, to obtain temporally aligned cross-modal features, we design a plug-and-play cross-modal interaction module between the encoder and decoder, seamlessly integrating visual and textual features. Comprehensive experiments on QVHighlights, Charades-STA, Activity-Net, and TVSum datasets show that MH-DETR outperforms existing state-of-the-art methods, demonstrating its effectiveness and superiority. Our code is available at https://github.com/YoucanBaby/MH-DETR.
Although lane detection methods have shown impressive performance in real-world scenarios, most of methods require post-processing which is not robust enough. Therefore, end-to-end detectors like DEtection TRansformer(DETR) have been introduced in lane detection.However, one-to-one label assignment in DETR can degrade the training efficiency due to label semantic conflicts. Besides, positional query in DETR is unable to provide explicit positional prior, making it difficult to be optimized. In this paper, we present the One-to-Several Transformer(O2SFormer). We first propose the one-to-several label assignment, which combines one-to-many and one-to-one label assignment to solve label semantic conflicts while keeping end-to-end detection. To overcome the difficulty in optimizing one-to-one assignment. We further propose the layer-wise soft label which dynamically adjusts the positive weight of positive lane anchors in different decoder layers. Finally, we design the dynamic anchor-based positional query to explore positional prior by incorporating lane anchors into positional query. Experimental results show that O2SFormer with ResNet50 backbone achieves 77.83% F1 score on CULane dataset, outperforming existing Transformer-based and CNN-based detectors. Futhermore, O2SFormer converges 12.5x faster than DETR for the ResNet18 backbone.
Live video commenting is popular on video media platforms, as it can create a chatting atmosphere and provide supplementary information for users while watching videos. Automatically generating live video comments can improve user experience and enable human-like generation for bot chatting. Existing works mostly focus on short video datasets while ignoring other important video types such as long videos like movies. In this work, we collect a new Movie Live Comments (MovieLC) dataset to support research on live video comment generation for long videos. We also propose a knowledge enhanced generation model inspired by the divergent and informative nature of live video comments. Our model adopts a pre-training encoder-decoder framework and incorporates external knowledge. Extensive experiments show that both objective metrics and human evaluation demonstrate the effectiveness of our proposed model. The MovieLC dataset and our code will be released.
Large language models (LLMs) with instruction fine-tuning demonstrate superior generative capabilities. However, these models are resource-intensive. To alleviate this issue, we explore distilling knowledge from instruction-tuned LLMs into much smaller ones. To this end, we carefully develop a large set of 2.58M instructions based on both existing and newly-generated instructions. In addition to being sizable, we design our instructions to cover a broad set of topics to ensure diversity. Extensive analysis of our instruction dataset confirms its diversity, and we generate responses for these instructions using gpt-3.5-turbo. Leveraging these instructions, we fine-tune a diverse herd of models, collectively referred to as LaMini-LM, which includes models from both the encoder-decoder and decoder-only families, with varying sizes. We evaluate the performance of our models using automatic metrics on 15 different natural language processing (NLP) benchmarks, as well as through human assessment. The results demonstrate that our proposed LaMini-LM models are comparable to competitive baselines, while being much smaller in size.
Neural fields are neural networks which map coordinates to a desired signal. When a neural field should jointly model multiple signals, and not memorize only one, it needs to be conditioned on a latent code which describes the signal at hand. Despite being an important aspect, there has been little research on conditioning strategies for neural fields. In this work, we explore the use of neural fields as decoders for 2D semantic segmentation. For this task, we compare three conditioning methods, simple concatenation of the latent code, Feature Wise Linear Modulation (FiLM), and Cross-Attention, in conjunction with latent codes which either describe the full image or only a local region of the image. Our results show a considerable difference in performance between the examined conditioning strategies. Furthermore, we show that conditioning via Cross-Attention achieves the best results and is competitive with a CNN-based decoder for semantic segmentation.
Keyphrase generation aims at generating topical phrases from a given text either by copying from the original text (present keyphrases) or by producing new keyphrases (absent keyphrases) that capture the semantic meaning of the text. Encoder-decoder models are most widely used for this task because of their capabilities for absent keyphrase generation. However, there has been little to no analysis on the performance and behavior of such models for keyphrase generation. In this paper, we study various tendencies exhibited by three strong models: T5 (based on a pre-trained transformer), CatSeq-Transformer (a non-pretrained Transformer), and ExHiRD (based on a recurrent neural network). We analyze prediction confidence scores, model calibration, and the effect of token position on keyphrases generation. Moreover, we motivate and propose a novel metric framework, SoftKeyScore, to evaluate the similarity between two sets of keyphrases by using softscores to account for partial matching and semantic similarity. We find that SoftKeyScore is more suitable than the standard F1 metric for evaluating two sets of given keyphrases.