Despite recent advances in modern machine learning algorithms, the opaqueness of their underlying mechanisms continues to be an obstacle in adoption. To instill confidence and trust in artificial intelligence systems, Explainable Artificial Intelligence has emerged as a response to improving modern machine learning algorithms' explainability. Inductive Logic Programming (ILP), a subfield of symbolic artificial intelligence, plays a promising role in generating interpretable explanations because of its intuitive logic-driven framework. ILP effectively leverages abductive reasoning to generate explainable first-order clausal theories from examples and background knowledge. However, several challenges in developing methods inspired by ILP need to be addressed for their successful application in practice. For example, existing ILP systems often have a vast solution space, and the induced solutions are very sensitive to noises and disturbances. This survey paper summarizes the recent advances in ILP and a discussion of statistical relational learning and neural-symbolic algorithms, which offer synergistic views to ILP. Following a critical review of the recent advances, we delineate observed challenges and highlight potential avenues of further ILP-motivated research toward developing self-explanatory artificial intelligence systems.
Recently, zero-shot image classification by vision-language pre-training has demonstrated incredible achievements, that the model can classify arbitrary category without seeing additional annotated images of that category. However, it is still unclear how to make the zero-shot recognition working well on broader vision problems, such as object detection and semantic segmentation. In this paper, we target for zero-shot semantic segmentation, by building it on an off-the-shelf pre-trained vision-language model, i.e., CLIP. It is difficult because semantic segmentation and the CLIP model perform on different visual granularity, that semantic segmentation processes on pixels while CLIP performs on images. To remedy the discrepancy on processing granularity, we refuse the use of the prevalent one-stage FCN based framework, and advocate a two-stage semantic segmentation framework, with the first stage extracting generalizable mask proposals and the second stage leveraging an image based CLIP model to perform zero-shot classification on the masked image crops which are generated in the first stage. Our experimental results show that this simple framework surpasses previous state-of-the-arts by a large margin: +29.5 hIoU on the Pascal VOC 2012 dataset, and +8.9 hIoU on the COCO Stuff dataset. With its simplicity and strong performance, we hope this framework to serve as a baseline to facilitate the future research.
This paper presents SimMIM, a simple framework for masked image modeling. We simplify recently proposed related approaches without special designs such as block-wise masking and tokenization via discrete VAE or clustering. To study what let the masked image modeling task learn good representations, we systematically study the major components in our framework, and find that simple designs of each component have revealed very strong representation learning performance: 1) random masking of the input image with a moderately large masked patch size (e.g., 32) makes a strong pre-text task; 2) predicting raw pixels of RGB values by direct regression performs no worse than the patch classification approaches with complex designs; 3) the prediction head can be as light as a linear layer, with no worse performance than heavier ones. Using ViT-B, our approach achieves 83.8% top-1 fine-tuning accuracy on ImageNet-1K by pre-training also on this dataset, surpassing previous best approach by +0.6%. When applied on a larger model of about 650 million parameters, SwinV2-H, it achieves 87.1% top-1 accuracy on ImageNet-1K using only ImageNet-1K data. We also leverage this approach to facilitate the training of a 3B model (SwinV2-G), that by $40\times$ less data than that in previous practice, we achieve the state-of-the-art on four representative vision benchmarks. The code and models will be publicly available at https://github.com/microsoft/SimMIM.
We present techniques for scaling Swin Transformer up to 3 billion parameters and making it capable of training with images of up to 1,536$\times$1,536 resolution. By scaling up capacity and resolution, Swin Transformer sets new records on four representative vision benchmarks: 84.0% top-1 accuracy on ImageNet-V2 image classification, 63.1/54.4 box/mask mAP on COCO object detection, 59.9 mIoU on ADE20K semantic segmentation, and 86.8% top-1 accuracy on Kinetics-400 video action classification. Our techniques are generally applicable for scaling up vision models, which has not been widely explored as that of NLP language models, partly due to the following difficulties in training and applications: 1) vision models often face instability issues at scale and 2) many downstream vision tasks require high resolution images or windows and it is not clear how to effectively transfer models pre-trained at low resolutions to higher resolution ones. The GPU memory consumption is also a problem when the image resolution is high. To address these issues, we present several techniques, which are illustrated by using Swin Transformer as a case study: 1) a post normalization technique and a scaled cosine attention approach to improve the stability of large vision models; 2) a log-spaced continuous position bias technique to effectively transfer models pre-trained at low-resolution images and windows to their higher-resolution counterparts. In addition, we share our crucial implementation details that lead to significant savings of GPU memory consumption and thus make it feasible to train large vision models with regular GPUs. Using these techniques and self-supervised pre-training, we successfully train a strong 3B Swin Transformer model and effectively transfer it to various vision tasks involving high-resolution images or windows, achieving the state-of-the-art accuracy on a variety of benchmarks.
We introduce MixTraining, a new training paradigm for object detection that can improve the performance of existing detectors for free. MixTraining enhances data augmentation by utilizing augmentations of different strengths while excluding the strong augmentations of certain training samples that may be detrimental to training. In addition, it addresses localization noise and missing labels in human annotations by incorporating pseudo boxes that can compensate for these errors. Both of these MixTraining capabilities are made possible through bootstrapping on the detector, which can be used to predict the difficulty of training on a strong augmentation, as well as to generate reliable pseudo boxes thanks to the robustness of neural networks to labeling error. MixTraining is found to bring consistent improvements across various detectors on the COCO dataset. In particular, the performance of Faster R-CNN \cite{ren2015faster} with a ResNet-50 \cite{he2016deep} backbone is improved from 41.7 mAP to 44.0 mAP, and the accuracy of Cascade-RCNN \cite{cai2018cascade} with a Swin-Small \cite{liu2021swin} backbone is raised from 50.9 mAP to 52.8 mAP. The code and models will be made publicly available at \url{https://github.com/MendelXu/MixTraining}.
Graph neural networks (GNNs) and label propagation represent two interrelated modeling strategies designed to exploit graph structure in tasks such as node property prediction. The former is typically based on stacked message-passing layers that share neighborhood information to transform node features into predictive embeddings. In contrast, the latter involves spreading label information to unlabeled nodes via a parameter-free diffusion process, but operates independently of the node features. Given then that the material difference is merely whether features or labels are smoothed across the graph, it is natural to consider combinations of the two for improving performance. In this regard, it has recently been proposed to use a randomly-selected portion of the training labels as GNN inputs, concatenated with the original node features for making predictions on the remaining labels. This so-called label trick accommodates the parallel use of features and labels, and is foundational to many of the top-ranking submissions on the Open Graph Benchmark (OGB) leaderboard. And yet despite its wide-spread adoption, thus far there has been little attempt to carefully unpack exactly what statistical properties the label trick introduces into the training pipeline, intended or otherwise. To this end, we prove that under certain simplifying assumptions, the stochastic label trick can be reduced to an interpretable, deterministic training objective composed of two factors. The first is a data-fitting term that naturally resolves potential label leakage issues, while the second serves as a regularization factor conditioned on graph structure that adapts to graph size and connectivity. Later, we leverage this perspective to motivate a broader range of label trick use cases, and provide experiments to verify the efficacy of these extensions.
Sarcasm employs ambivalence, where one says something positive but actually means negative, and vice versa. Due to the sophisticated and obscure sentiment, sarcasm brings in great challenges to sentiment analysis. In this paper, we show up the essence of sarcastic text is that the literal sentiment (expressed by the surface form of the text) is opposite to the deep sentiment (expressed by the actual meaning of the text). To this end, we propose a Dual-Channel Framework by modeling both literal and deep sentiments to recognize the sentiment conflict. Specifically, the proposed framework is capable of detecting the sentiment conflict between the literal and deep meanings of the input text. Experiments on the political debates and the Twitter datasets show that our framework achieves the best performance on sarcasm recognition.
Kinship verification from facial images has been recognized as an emerging yet challenging technique in many potential computer vision applications. In this paper, we propose a novel cross-generation feature interaction learning (CFIL) framework for robust kinship verification. Particularly, an effective collaborative weighting strategy is constructed to explore the characteristics of cross-generation relations by corporately extracting features of both parents and children image pairs. Specifically, we take parents and children as a whole to extract the expressive local and non-local features. Different from the traditional works measuring similarity by distance, we interpolate the similarity calculations as the interior auxiliary weights into the deep CNN architecture to learn the whole and natural features. These similarity weights not only involve corresponding single points but also excavate the multiple relationships cross points, where local and non-local features are calculated by using these two kinds of distance measurements. Importantly, instead of separately conducting similarity computation and feature extraction, we integrate similarity learning and feature extraction into one unified learning process. The integrated representations deduced from local and non-local features can comprehensively express the informative semantics embedded in images and preserve abundant correlation knowledge from image pairs. Extensive experiments demonstrate the efficiency and superiority of the proposed model compared to some state-of-the-art kinship verification methods.
Kinship verification is a long-standing research challenge in computer vision. The visual differences presented to the face have a significant effect on the recognition capabilities of the kinship systems. We argue that aggregating multiple visual knowledge can better describe the characteristics of the subject for precise kinship identification. Typically, the age-invariant features can represent more natural facial details. Such age-related transformations are essential for face recognition due to the biological effects of aging. However, the existing methods mainly focus on employing the single-view image features for kinship identification, while more meaningful visual properties such as race and age are directly ignored in the feature learning step. To this end, we propose a novel deep collaborative multi-modal learning (DCML) to integrate the underlying information presented in facial properties in an adaptive manner to strengthen the facial details for effective unsupervised kinship verification. Specifically, we construct a well-designed adaptive feature fusion mechanism, which can jointly leverage the complementary properties from different visual perspectives to produce composite features and draw greater attention to the most informative components of spatial feature maps. Particularly, an adaptive weighting strategy is developed based on a novel attention mechanism, which can enhance the dependencies between different properties by decreasing the information redundancy in channels in a self-adaptive manner. To validate the effectiveness of the proposed method, extensive experimental evaluations conducted on four widely-used datasets show that our DCML method is always superior to some state-of-the-art kinship verification methods.