Vision-Language Pre-Training (VLP) has shown promising capabilities to align image and text pairs, facilitating a broad variety of cross-modal learning tasks. However, we observe that VLP models often lack the visual grounding/localization capability which is critical for many downstream tasks such as visual reasoning. In this work, we propose a novel Position-guided Text Prompt (PTP) paradigm to enhance the visual grounding ability of cross-modal models trained with VLP. Specifically, in the VLP phase, PTP divides the image into $N\times N$ blocks, and identifies the objects in each block through the widely used object detector in VLP. It then reformulates the visual grounding task into a fill-in-the-blank problem given a PTP by encouraging the model to predict the objects in the given blocks or regress the blocks of a given object, e.g. filling `P" or ``O" in aPTP ``The block P has a O". This mechanism improves the visual grounding capability of VLP models and thus helps them better handle various downstream tasks. By introducing PTP into several state-of-the-art VLP frameworks, we observe consistently significant improvements across representative cross-modal learning model architectures and several benchmarks, e.g. zero-shot Flickr30K Retrieval (+4.8 in average recall@1) for ViLT \cite{vilt} baseline, and COCO Captioning (+5.3 in CIDEr) for SOTA BLIP \cite{blip} baseline. Moreover, PTP achieves comparable results with object-detector based methods, and much faster inference speed since PTP discards its object detector for inference while the later cannot. Our code and pre-trained weight will be released at \url{https://github.com/sail-sg/ptp}.
Directly training a document-to-document (Doc2Doc) neural machine translation (NMT) via Transformer from scratch, especially on small datasets usually fails to converge. Our dedicated probing tasks show that 1) both the absolute position and relative position information gets gradually weakened or even vanished once it reaches the upper encoder layers, and 2) the vanishing of absolute position information in encoder output causes the training failure of Doc2Doc NMT. To alleviate this problem, we propose a position-aware Transformer (P-Transformer) to enhance both the absolute and relative position information in both self-attention and cross-attention. Specifically, we integrate absolute positional information, i.e., position embeddings, into the query-key pairs both in self-attention and cross-attention through a simple yet effective addition operation. Moreover, we also integrate relative position encoding in self-attention. The proposed P-Transformer utilizes sinusoidal position encoding and does not require any task-specified position embedding, segment embedding, or attention mechanism. Through the above methods, we build a Doc2Doc NMT model with P-Transformer, which ingests the source document and completely generates the target document in a sequence-to-sequence (seq2seq) way. In addition, P-Transformer can be applied to seq2seq-based document-to-sentence (Doc2Sent) and sentence-to-sentence (Sent2Sent) translation. Extensive experimental results of Doc2Doc NMT show that P-Transformer significantly outperforms strong baselines on widely-used 9 document-level datasets in 7 language pairs, covering small-, middle-, and large-scales, and achieves a new state-of-the-art. Experimentation on discourse phenomena shows that our Doc2Doc NMT models improve the translation quality in both BLEU and discourse coherence. We make our code available on Github.
The Position Embedding (PE) is critical for Vision Transformers (VTs) due to the permutation-invariance of self-attention operation. By analyzing the input and output of each encoder layer in VTs using reparameterization and visualization, we find that the default PE joining method (simply adding the PE and patch embedding together) operates the same affine transformation to token embedding and PE, which limits the expressiveness of PE and hence constrains the performance of VTs. To overcome this limitation, we propose a simple, effective, and robust method. Specifically, we provide two independent layer normalizations for token embeddings and PE for each layer, and add them together as the input of each layer's Muti-Head Self-Attention module. Since the method allows the model to adaptively adjust the information of PE for different layers, we name it as Layer-adaptive Position Embedding, abbreviated as LaPE. Extensive experiments demonstrate that LaPE can improve various VTs with different types of PE and make VTs robust to PE types. For example, LaPE improves 0.94% accuracy for ViT-Lite on Cifar10, 0.98% for CCT on Cifar100, and 1.72% for DeiT on ImageNet-1K, which is remarkable considering the negligible extra parameters, memory and computational cost brought by LaPE. The code is publicly available at https://github.com/Ingrid725/LaPE.
This paper proposes a Learnable Multiplicative absolute position Embedding based Conformer (LMEC). It contains a kernelized linear attention (LA) module called LMLA to solve the time-consuming problem for long sequence speech recognition as well as an alternative to the FFN structure. First, the ELU function is adopted as the kernel function of our proposed LA module. Second, we propose a novel Learnable Multiplicative Absolute Position Embedding (LM-APE) based re-weighting mechanism that can reduce the well-known quadratic temporal-space complexity of softmax self-attention. Third, we use Gated Linear Units (GLU) to substitute the Feed Forward Network (FFN) for better performance. Extensive experiments have been conducted on the public LibriSpeech datasets. Compared to the Conformer model with cosFormer style linear attention, our proposed method can achieve up to 0.63% word-error-rate improvement on test-other and improve the inference speed by up to 13% (left product) and 33% (right product) on the LA module.
Motivation: Enhancers are important cis-regulatory elements that regulate a wide range of biological functions and enhance the transcription of target genes. Although many feature extraction methods have been proposed to improve the performance of enhancer identification, they cannot learn position-related multiscale contextual information from raw DNA sequences. Results: In this article, we propose a novel enhancer identification method (iEnhancer-ELM) based on BERT-like enhancer language models. iEnhancer-ELM tokenizes DNA sequences with multi-scale k-mers and extracts contextual information of different scale k-mers related with their positions via an multi-head attention mechanism. We first evaluate the performance of different scale k-mers, then ensemble them to improve the performance of enhancer identification. The experimental results on two popular benchmark datasets show that our model outperforms stateof-the-art methods. We further illustrate the interpretability of iEnhancer-ELM. For a case study, we discover 30 enhancer motifs via a 3-mer-based model, where 12 of motifs are verified by STREME and JASPAR, demonstrating our model has a potential ability to unveil the biological mechanism of enhancer. Availability and implementation: The models and associated code are available at https://github.com/chen-bioinfo/iEnhancer-ELM Contact: [email omitted] Supplementary information: Supplementary data are available at Bioinformatics Advances online.
Graph generative models have broad applications in biology, chemistry and social science. However, modelling and understanding the generative process of graphs is challenging due to the discrete and high-dimensional nature of graphs, as well as permutation invariance to node orderings in underlying graph distributions. Current leading autoregressive models fail to capture the permutation invariance nature of graphs for the reliance on generation ordering and have high time complexity. Here, we propose a continuous-time generative diffusion process for permutation invariant graph generation to mitigate these issues. Specifically, we first construct a forward diffusion process defined by a stochastic differential equation (SDE), which smoothly converts graphs within the complex distribution to random graphs that follow a known edge probability. Solving the corresponding reverse-time SDE, graphs can be generated from newly sampled random graphs. To facilitate the reverse-time SDE, we newly design a position-enhanced graph score network, capturing the evolving structure and position information from perturbed graphs for permutation equivariant score estimation. Under the evaluation of comprehensive metrics, our proposed generative diffusion process achieves competitive performance in graph distribution learning. Experimental results also show that GraphGDP can generate high-quality graphs in only 24 function evaluations, much faster than previous autoregressive models.
The introductory programming sequence has been the focus of much research in computing education. The recent advent of several viable and freely-available AI-driven code generation tools present several immediate opportunities and challenges in this domain. In this position paper we argue that the community needs to act quickly in deciding what possible opportunities can and should be leveraged and how, while also working on how to overcome or otherwise mitigate the possible challenges. Assuming that the effectiveness and proliferation of these tools will continue to progress rapidly, without quick, deliberate, and concerted efforts, educators will lose advantage in helping shape what opportunities come to be, and what challenges will endure. With this paper we aim to seed this discussion within the computing education community.
Incorporating prior knowledge of physics laws and structural properties of dynamical systems into the design of deep learning architectures has proven to be a powerful technique for improving their computational efficiency and generalization capacity. Learning accurate models of robot dynamics is critical for safe and stable control. Autonomous mobile robots, including wheeled, aerial, and underwater vehicles, can be modeled as controlled Lagrangian or Hamiltonian rigid-body systems evolving on matrix Lie groups. In this paper, we introduce a new structure-preserving deep learning architecture, the Lie group Forced Variational Integrator Network (LieFVIN), capable of learning controlled Lagrangian or Hamiltonian dynamics on Lie groups, either from position-velocity or position-only data. By design, LieFVINs preserve both the Lie group structure on which the dynamics evolve and the symplectic structure underlying the Hamiltonian or Lagrangian systems of interest. The proposed architecture learns surrogate discrete-time flow maps allowing accurate and fast prediction without numerical-integrator, neural-ODE, or adjoint techniques, which are needed for vector fields. Furthermore, the learnt discrete-time dynamics can be utilized with computationally scalable discrete-time (optimal) control strategies.
Automatic diabetic retinopathy (DR) grading based on fundus photography has been widely explored to benefit the routine screening and early treatment. Existing researches generally focus on single-field fundus images, which have limited field of view for precise eye examinations. In clinical applications, ophthalmologists adopt two-field fundus photography as the dominating tool, where the information from each field (i.e.,macula-centric and optic disc-centric) is highly correlated and complementary, and benefits comprehensive decisions. However, automatic DR grading based on two-field fundus photography remains a challenging task due to the lack of publicly available datasets and effective fusion strategies. In this work, we first construct a new benchmark dataset (DRTiD) for DR grading, consisting of 3,100 two-field fundus images. To the best of our knowledge, it is the largest public DR dataset with diverse and high-quality two-field images. Then, we propose a novel DR grading approach, namely Cross-Field Transformer (CrossFiT), to capture the correspondence between two fields as well as the long-range spatial correlations within each field. Considering the inherent two-field geometric constraints, we particularly define aligned position embeddings to preserve relative consistent position in fundus. Besides, we perform masked cross-field attention during interaction to flter the noisy relations between fields. Extensive experiments on our DRTiD dataset and a public DeepDRiD dataset demonstrate the effectiveness of our CrossFiT network. The new dataset and the source code of CrossFiT will be publicly available at https://github.com/FDU-VTS/DRTiD.
Because anomalous samples cannot be used for training, many anomaly detection and localization methods use pre-trained networks and non-parametric modeling to estimate encoded feature distribution. However, these methods neglect the impact of position and neighborhood information on the distribution of normal features. To overcome this, we propose a new algorithm, \textbf{PNI}, which estimates the normal distribution using conditional probability given neighborhood features, modeled with a multi-layer perceptron network. Moreover, position information is utilized by creating a histogram of representative features at each position. Instead of simply resizing the anomaly map, the proposed method employs an additional refine network trained on synthetic anomaly images to better interpolate and account for the shape and edge of the input image. We conducted experiments on the MVTec AD benchmark dataset and achieved state-of-the-art performance, with \textbf{99.56\%} and \textbf{98.98\%} AUROC scores in anomaly detection and localization, respectively.
Despite impressive performance for high-level downstream tasks, self-supervised pre-training methods have not yet fully delivered on dense geometric vision tasks such as stereo matching or optical flow. The application of self-supervised concepts, such as instance discrimination or masked image modeling, to geometric tasks is an active area of research. In this work, we build on the recent cross-view completion framework, a variation of masked image modeling that leverages a second view from the same scene which makes it well suited for binocular downstream tasks. The applicability of this concept has so far been limited in at least two ways: (a) by the difficulty of collecting real-world image pairs -- in practice only synthetic data have been used -- and (b) by the lack of generalization of vanilla transformers to dense downstream tasks for which relative position is more meaningful than absolute position. We explore three avenues of improvement. First, we introduce a method to collect suitable real-world image pairs at large scale. Second, we experiment with relative positional embeddings and show that they enable vision transformers to perform substantially better. Third, we scale up vision transformer based cross-completion architectures, which is made possible by the use of large amounts of data. With these improvements, we show for the first time that state-of-the-art results on stereo matching and optical flow can be reached without using any classical task-specific techniques like correlation volume, iterative estimation, image warping or multi-scale reasoning, thus paving the way towards universal vision models.
Wave propagation through nodes and links of a network forms the basis of spectral graph theory. Nevertheless, the sound emitted by nodes within the resonating chamber formed by a network are not well studied. The sound emitted by vibrations of individual nodes reflects the structure of the overall network topology but also the location of the node within the network. In this article, a sound recognition neural network is trained to infer centrality measures from the nodes' wave-forms. In addition to advancing network representation learning, sounds emitted by nodes are plausible in most cases. Auralization of the network topology may open new directions in arts, competing with network visualization.
Autonomous underwater vehicles (AUVs) are becoming standard tools for underwater exploration and seabed mapping in both scientific and industrial applications \cite{graham2022rapid, stenius2022system}. Their capacity to dive untethered allows them to reach areas inaccessible to surface vessels and to collect data more closely to the seafloor, regardless of the water depth. However, their navigation autonomy remains bounded by the accuracy of their dead reckoning (DR) estimate of their global position, severely limited in the absence of a priori maps of the area and GPS signal. Global localization systems equivalent to the later exists for the underwater domain, such as LBL or USBL. However they involve expensive external infrastructure and their reliability decreases with the distance to the AUV, making them unsuitable for deep sea surveys.
Rejecting correspondence outliers enables to boost the correspondence quality, which is a critical step in achieving high point cloud registration accuracy. The current state-of-the-art correspondence outlier rejection methods only utilize the structure features of the correspondences. However, texture information is critical to reject the correspondence outliers in our human vision system. In this paper, we propose General Multimodal Fusion (GMF) to learn to reject the correspondence outliers by leveraging both the structure and texture information. Specifically, two cross-attention-based fusion layers are proposed to fuse the texture information from paired images and structure information from point correspondences. Moreover, we propose a convolutional position encoding layer to enhance the difference between Tokens and enable the encoding feature pay attention to neighbor information. Our position encoding layer will make the cross-attention operation integrate both local and global information. Experiments on multiple datasets(3DMatch, 3DLoMatch, KITTI) and recent state-of-the-art models (3DRegNet, DGR, PointDSC) prove that our GMF achieves wide generalization ability and consistently improves the point cloud registration accuracy. Furthermore, several ablation studies demonstrate the robustness of the proposed GMF on different loss functions, lighting conditions and noises.The code is available at https://github.com/XiaoshuiHuang/GMF.
Data-efficient learning on graphs (GEL) is essential in real-world applications. Existing GEL methods focus on learning useful representations for nodes, edges, or entire graphs with ``small'' labeled data. But the problem of data-efficient learning for subgraph prediction has not been explored. The challenges of this problem lie in the following aspects: 1) It is crucial for subgraphs to learn positional features to acquire structural information in the base graph in which they exist. Although the existing subgraph neural network method is capable of learning disentangled position encodings, the overall computational complexity is very high. 2) Prevailing graph augmentation methods for GEL, including rule-based, sample-based, adaptive, and automated methods, are not suitable for augmenting subgraphs because a subgraph contains fewer nodes but richer information such as position, neighbor, and structure. Subgraph augmentation is more susceptible to undesirable perturbations. 3) Only a small number of nodes in the base graph are contained in subgraphs, which leads to a potential ``bias'' problem that the subgraph representation learning is dominated by these ``hot'' nodes. By contrast, the remaining nodes fail to be fully learned, which reduces the generalization ability of subgraph representation learning. In this paper, we aim to address the challenges above and propose a Position-Aware Data-Efficient Learning framework for subgraph neural networks called PADEL. Specifically, we propose a novel node position encoding method that is anchor-free, and design a new generative subgraph augmentation method based on a diffused variational subgraph autoencoder, and we propose exploratory and exploitable views for subgraph contrastive learning. Extensive experiment results on three real-world datasets show the superiority of our proposed method over state-of-the-art baselines.
In this paper, we present our approach and empirical observations for Cause-Effect Signal Span Detection -- Subtask 2 of Shared task 3~\cite{tan-etal-2022-event} at CASE 2022. The shared task aims to extract the cause, effect, and signal spans from a given causal sentence. We model the task as a reading comprehension (RC) problem and apply a token-level RC-based span prediction paradigm to the task as the baseline. We explore different training objectives to fine-tune the model, as well as data augmentation (DA) tricks based on the language model (LM) for performance improvement. Additionally, we propose an efficient beam-search post-processing strategy to due with the drawbacks of span detection to obtain a further performance gain. Our approach achieves an average $F_1$ score of 54.15 and ranks \textbf{$1^{st}$} in the CASE competition. Our code is available at \url{https://github.com/Gzhang-umich/1CademyTeamOfCASE}.
Inference is a significant part of ML software infrastructure. Despite the variety of inference frameworks available, the field as a whole can be considered in its early days. This position paper puts forth a range of important qualities that next generation of inference platforms should be aiming for. We present our rationale for the importance of each quality, and discuss ways to achieve it in practice. We propose to focus on data-centricity as the overarching design pattern which enables smarter ML system deployment and operation at scale.
Transformer language models encode the notion of word order using positional information. Most commonly, this positional information is represented by absolute position embeddings (APEs), that are learned from the pretraining data. However, in natural language, it is not absolute position that matters, but relative position, and the extent to which APEs can capture this type of information has not been investigated. In this work, we observe that models trained with APE over-rely on positional information to the point that they break-down when subjected to sentences with shifted position information. Specifically, when models are subjected to sentences starting from a non-zero position (excluding the effect of priming), they exhibit noticeably degraded performance on zero to full-shot tasks, across a range of model families and model sizes. Our findings raise questions about the efficacy of APEs to model the relativity of position information, and invite further introspection on the sentence and word order processing strategies employed by these models.
Cujo AI and Adversa AI hosted the MLSec face recognition challenge. The goal was to attack a black box face recognition model with targeted attacks. The model returned the confidence of the target class and a stealthiness score. For an attack to be considered successful the target class has to have the highest confidence among all classes and the stealthiness has to be at least 0.5. In our approach we paste the face of a target into a source image. By utilizing position, scaling, rotation and transparency attributes we reached 3rd place. Our approach took approximately 200 queries per attack for the final highest score and about ~7.7 queries minimum for a successful attack. The code is available at https://github.com/bunni90/FacePastingAttack .
We introduce MOVE, a novel method to segment objects without any form of supervision. MOVE exploits the fact that foreground objects can be shifted locally relative to their initial position and result in realistic (undistorted) new images. This property allows us to train a segmentation model on a dataset of images without annotation and to achieve state of the art (SotA) performance on several evaluation datasets for unsupervised salient object detection and segmentation. In unsupervised single object discovery, MOVE gives an average CorLoc improvement of 7.2% over the SotA, and in unsupervised class-agnostic object detection it gives a relative AP improvement of 53% on average. Our approach is built on top of self-supervised features (e.g. from DINO or MAE), an inpainting network (based on the Masked AutoEncoder) and adversarial training.