Many experts argue that the future of artificial intelligence is limited by the field's ability to integrate symbolic logical reasoning into deep learning architectures. The recently proposed differentiable MAXSAT solver, SATNet, was a breakthrough in its capacity to integrate with a traditional neural network and solve visual reasoning problems. For instance, it can learn the rules of Sudoku purely from image examples. Despite its success, SATNet was shown to succumb to a key challenge in neurosymbolic systems known as the Symbol Grounding Problem: the inability to map visual inputs to symbolic variables without explicit supervision ("label leakage"). In this work, we present a self-supervised pre-training pipeline that enables SATNet to overcome this limitation, thus broadening the class of problems that SATNet architectures can solve to include datasets where no intermediary labels are available at all. We demonstrate that our method allows SATNet to attain full accuracy even with a harder problem setup that prevents any label leakage. We additionally introduce a proofreading method that further improves the performance of SATNet architectures, beating the state-of-the-art on Visual Sudoku.
Image captioning is an important task for benchmarking visual reasoning and for enabling accessibility for people with vision impairments. However, as in many machine learning settings, social biases can influence image captioning in undesirable ways. In this work, we study bias propagation pathways within image captioning, focusing specifically on the COCO dataset. Prior work has analyzed gender bias in captions using automatically-derived gender labels; here we examine racial and intersectional biases using manual annotations. Our first contribution is in annotating the perceived gender and skin color of 28,315 of the depicted people after obtaining IRB approval. Using these annotations, we compare racial biases present in both manual and automatically-generated image captions. We demonstrate differences in caption performance, sentiment, and word choice between images of lighter versus darker-skinned people. Further, we find the magnitude of these differences to be greater in modern captioning systems compared to older ones, thus leading to concerns that without proper consideration and mitigation these differences will only become increasingly prevalent. Code and data is available at https://princetonvisualai.github.io/imagecaptioning-bias .
As an important step towards visual reasoning, visual grounding (e.g., phrase localization, referring expression comprehension/segmentation) has been widely explored Previous approaches to referring expression comprehension (REC) or segmentation (RES) either suffer from limited performance, due to a two-stage setup, or require the designing of complex task-specific one-stage architectures. In this paper, we propose a simple one-stage multi-task framework for visual grounding tasks. Specifically, we leverage a transformer architecture, where two modalities are fused in a visual-lingual encoder. In the decoder, the model learns to generate contextualized lingual queries which are then decoded and used to directly regress the bounding box and produce a segmentation mask for the corresponding referred regions. With this simple but highly contextualized model, we outperform state-of-the-arts methods by a large margin on both REC and RES tasks. We also show that a simple pre-training schedule (on an external dataset) further improves the performance. Extensive experiments and ablations illustrate that our model benefits greatly from contextualized information and multi-task training.
Despite the achievements of large-scale multimodal pre-training approaches, cross-modal retrieval, e.g., image-text retrieval, remains a challenging task. To bridge the semantic gap between the two modalities, previous studies mainly focus on word-region alignment at the object level, lacking the matching between the linguistic relation among the words and the visual relation among the regions. The neglect of such relation consistency impairs the contextualized representation of image-text pairs and hinders the model performance and the interpretability. In this paper, we first propose a novel metric, Intra-modal Self-attention Distance (ISD), to quantify the relation consistency by measuring the semantic distance between linguistic and visual relations. In response, we present Inter-modal Alignment on Intra-modal Self-attentions (IAIS), a regularized training method to optimize the ISD and calibrate intra-modal self-attentions from the two modalities mutually via inter-modal alignment. The IAIS regularizer boosts the performance of prevailing models on Flickr30k and MS COCO datasets by a considerable margin, which demonstrates the superiority of our approach.
Geometry problem solving has attracted much attention in the NLP community recently. The task is challenging as it requires abstract problem understanding and symbolic reasoning with axiomatic knowledge. However, current datasets are either small in scale or not publicly available. Thus, we construct a new large-scale benchmark, Geometry3K, consisting of 3,002 geometry problems with dense annotation in formal language. We further propose a novel geometry solving approach with formal language and symbolic reasoning, called Interpretable Geometry Problem Solver (Inter-GPS). Inter-GPS first parses the problem text and diagram into formal language automatically via rule-based text parsing and neural object detecting, respectively. Unlike implicit learning in existing methods, Inter-GPS incorporates theorem knowledge as conditional rules and performs symbolic reasoning step by step. Also, a theorem predictor is designed to infer the theorem application sequence fed to the symbolic solver for the more efficient and reasonable searching path. Extensive experiments on the Geometry3K and GEOS datasets demonstrate that Inter-GPS achieves significant improvements over existing methods. The project with code and data is available at https://lupantech.github.io/inter-gps.
Referring Expression Comprehension (REC) has become one of the most important tasks in visual reasoning, since it is an essential step for many vision-and-language tasks such as visual question answering. However, it has not been widely used in many downstream tasks because it suffers 1) two-stage methods exist heavy computation cost and inevitable error accumulation, and 2) one-stage methods have to depend on lots of hyper-parameters (such as anchors) to generate bounding box. In this paper, we present a proposal-free one-stage (PFOS) model that is able to regress the region-of-interest from the image, based on a textual query, in an end-to-end manner. Instead of using the dominant anchor proposal fashion, we directly take the dense-grid of an image as input for a cross-attention transformer that learns grid-word correspondences. The final bounding box is predicted directly from the image without the time-consuming anchor selection process that previous methods suffer. Our model achieves the state-of-the-art performance on four referring expression datasets with higher efficiency, comparing to previous best one-stage and two-stage methods.
Vision-and-Language (VL) pre-training has shown great potential on many related downstream tasks, such as Visual Question Answering (VQA), one of the most popular problems in the VL field. All of these pre-trained models (such as VisualBERT, ViLBERT, LXMERT and UNITER) are built with Transformer, which extends the classical attention mechanism to multiple layers and heads. To investigate why and how these models work on VQA so well, in this paper we explore the roles of individual heads and layers in Transformer models when handling $12$ different types of questions. Specifically, we manually remove (chop) heads (or layers) from a pre-trained VisualBERT model at a time, and test it on different levels of questions to record its performance. As shown in the interesting echelon shape of the result matrices, experiments reveal different heads and layers are responsible for different question types, with higher-level layers activated by higher-level visual reasoning questions. Based on this observation, we design a dynamic chopping module that can automatically remove heads and layers of the VisualBERT at an instance level when dealing with different questions. Our dynamic chopping module can effectively reduce the parameters of the original model by 50%, while only damaging the accuracy by less than 1% on the VQA task.
Recent advances in natural language processing and computer vision have led to AI models that interpret simple scenes at human levels. Yet, we do not have a complete understanding of how humans and AI models differ in their interpretation of more complex scenes. We created a dataset of complex scenes that contained human behaviors and social interactions. AI and humans had to describe the scenes with a sentence. We used a quantitative metric of similarity between scene descriptions of the AI/human and ground truth of five other human descriptions of each scene. Results show that the machine/human agreement scene descriptions are much lower than human/human agreement for our complex scenes. Using an experimental manipulation that occludes different spatial regions of the scenes, we assessed how machines and humans vary in utilizing regions of images to understand the scenes. Together, our results are a first step toward understanding how machines fall short of human visual reasoning with complex scenes depicting human behaviors.
We study joint learning of Convolutional Neural Network (CNN) and Transformer for vision-language pre-training (VLPT) which aims to learn cross-modal alignments from millions of image-text pairs. State-of-the-art approaches extract salient image regions and align regions with words step-by-step. As region-based visual features usually represent parts of an image, it is challenging for existing vision-language models to fully understand the semantics from paired natural languages. In this paper, we propose SOHO to "See Out of tHe bOx" that takes a whole image as input, and learns vision-language representation in an end-to-end manner. SOHO does not require bounding box annotations which enables inference 10 times faster than region-based approaches. In particular, SOHO learns to extract comprehensive yet compact image features through a visual dictionary (VD) that facilitates cross-modal understanding. VD is designed to represent consistent visual abstractions of similar semantics. It is updated on-the-fly and utilized in our proposed pre-training task Masked Visual Modeling (MVM). We conduct experiments on four well-established vision-language tasks by following standard VLPT settings. In particular, SOHO achieves absolute gains of 2.0% R@1 score on MSCOCO text retrieval 5k test split, 1.5% accuracy on NLVR$^2$ test-P split, 6.7% accuracy on SNLI-VE test split, respectively.
Visual events are a composition of temporal actions involving actors spatially interacting with objects. When developing computer vision models that can reason about compositional spatio-temporal events, we need benchmarks that can analyze progress and uncover shortcomings. Existing video question answering benchmarks are useful, but they often conflate multiple sources of error into one accuracy metric and have strong biases that models can exploit, making it difficult to pinpoint model weaknesses. We present Action Genome Question Answering (AGQA), a new benchmark for compositional spatio-temporal reasoning. AGQA contains $192M$ unbalanced question answer pairs for $9.6K$ videos. We also provide a balanced subset of $3.9M$ question answer pairs, $3$ orders of magnitude larger than existing benchmarks, that minimizes bias by balancing the answer distributions and types of question structures. Although human evaluators marked $86.02\%$ of our question-answer pairs as correct, the best model achieves only $47.74\%$ accuracy. In addition, AGQA introduces multiple training/test splits to test for various reasoning abilities, including generalization to novel compositions, to indirect references, and to more compositional steps. Using AGQA, we evaluate modern visual reasoning systems, demonstrating that the best models barely perform better than non-visual baselines exploiting linguistic biases and that none of the existing models generalize to novel compositions unseen during training.
We study the problem of dynamic visual reasoning on raw videos. This is a challenging problem; currently, state-of-the-art models often require dense supervision on physical object properties and events from simulation, which are impractical to obtain in real life. In this paper, we present the Dynamic Concept Learner (DCL), a unified framework that grounds physical objects and events from video and language. DCL first adopts a trajectory extractor to track each object over time and to represent it as a latent, object-centric feature vector. Building upon this object-centric representation, DCL learns to approximate the dynamic interaction among objects using graph networks. DCL further incorporates a semantic parser to parse questions into semantic programs and, finally, a program executor to run the program to answer the question, levering the learned dynamics model. After training, DCL can detect and associate objects across the frames, ground visual properties, and physical events, understand the causal relationship between events, make future and counterfactual predictions, and leverage these extracted presentations for answering queries. DCL achieves state-of-the-art performance on CLEVRER, a challenging causal video reasoning dataset, even without using ground-truth attributes and collision labels from simulations for training. We further test DCL on a newly proposed video-retrieval and event localization dataset derived from CLEVRER, showing its strong generalization capacity.
GQA~\citep{hudson2019gqa} is a dataset for real-world visual reasoning and compositional question answering. We found that many answers predicted by the best vision-language models on the GQA dataset do not match the ground-truth answer but still are semantically meaningful and correct in the given context. In fact, this is the case with most existing visual question answering (VQA) datasets where they assume only one ground-truth answer for each question. We propose Alternative Answer Sets (AAS) of ground-truth answers to address this limitation, which is created automatically using off-the-shelf NLP tools. We introduce a semantic metric based on AAS and modify top VQA solvers to support multiple plausible answers for a question. We implement this approach on the GQA dataset and show the performance improvements. Code and data are available in this link \url{https://github.com/luomancs/alternative_answer_set.git}.
Causal induction, i.e., identifying unobservable mechanisms that lead to the observable relations among variables, has played a pivotal role in modern scientific discovery, especially in scenarios with only sparse and limited data. Humans, even young toddlers, can induce causal relationships surprisingly well in various settings despite its notorious difficulty. However, in contrast to the commonplace trait of human cognition is the lack of a diagnostic benchmark to measure causal induction for modern Artificial Intelligence (AI) systems. Therefore, in this work, we introduce the Abstract Causal REasoning (ACRE) dataset for systematic evaluation of current vision systems in causal induction. Motivated by the stream of research on causal discovery in Blicket experiments, we query a visual reasoning system with the following four types of questions in either an independent scenario or an interventional scenario: direct, indirect, screening-off, and backward-blocking, intentionally going beyond the simple strategy of inducing causal relationships by covariation. By analyzing visual reasoning architectures on this testbed, we notice that pure neural models tend towards an associative strategy under their chance-level performance, whereas neuro-symbolic combinations struggle in backward-blocking reasoning. These deficiencies call for future research in models with a more comprehensive capability of causal induction.
Abstract reasoning ability is fundamental to human intelligence. It enables humans to uncover relations among abstract concepts and further deduce implicit rules from the relations. As a well-known abstract visual reasoning task, Raven's Progressive Matrices (RPM) are widely used in human IQ tests. Although extensive research has been conducted on RPM solvers with machine intelligence, few studies have considered further advancing the standard answer-selection (classification) problem to a more challenging answer-painting (generating) problem, which can verify whether the model has indeed understood the implicit rules. In this paper we aim to solve the latter one by proposing a deep latent variable model, in which multiple Gaussian processes are employed as priors of latent variables to separately learn underlying abstract concepts from RPMs; thus the proposed model is interpretable in terms of concept-specific latent variables. The latent Gaussian process also provides an effective way of extrapolation for answer painting based on the learned concept-changing rules. We evaluate the proposed model on RPM-like datasets with multiple continuously-changing visual concepts. Experimental results demonstrate that our model requires only few training samples to paint high-quality answers, generate novel RPM panels, and achieve interpretability through concept-specific latent variables.
Raven's Progressive Matrices (RPMs) are frequently used in testing human's visual reasoning ability. Recent advances of RPM-like datasets and solution models partially address the challenges of visually understanding the RPM questions and logically reasoning the missing answers. In view of the poor generalization performance due to insufficient samples in RPM datasets, we propose an effective scheme, namely Candidate Answer Morphological Mixup (CAM-Mix). CAM-Mix serves as a data augmentation strategy by gray-scale image morphological mixup, which regularizes various solution methods and overcomes the model overfitting problem. By creating new negative candidate answers semantically similar to the correct answers, a more accurate decision boundary could be defined. By applying the proposed data augmentation method, a significant and consistent performance improvement is achieved on various RPM-like datasets compared with the state-of-the-art models.
State-of-the-art computer vision systems are trained to predict a fixed set of predetermined object categories. This restricted form of supervision limits their generality and usability since additional labeled data is needed to specify any other visual concept. Learning directly from raw text about images is a promising alternative which leverages a much broader source of supervision. We demonstrate that the simple pre-training task of predicting which caption goes with which image is an efficient and scalable way to learn SOTA image representations from scratch on a dataset of 400 million (image, text) pairs collected from the internet. After pre-training, natural language is used to reference learned visual concepts (or describe new ones) enabling zero-shot transfer of the model to downstream tasks. We study the performance of this approach by benchmarking on over 30 different existing computer vision datasets, spanning tasks such as OCR, action recognition in videos, geo-localization, and many types of fine-grained object classification. The model transfers non-trivially to most tasks and is often competitive with a fully supervised baseline without the need for any dataset specific training. For instance, we match the accuracy of the original ResNet-50 on ImageNet zero-shot without needing to use any of the 1.28 million training examples it was trained on. We release our code and pre-trained model weights at https://github.com/OpenAI/CLIP.
A common approach to solving physical reasoning tasks is to train a value learner on example tasks. A limitation of such an approach is that it requires learning about object dynamics solely from reward values assigned to the final state of a rollout of the environment. This study aims to address this limitation by augmenting the reward value with self-supervised signals about object dynamics. Specifically, we train the model to characterize the similarity of two environment rollouts, jointly with predicting the outcome of the reasoning task. This similarity can be defined as a distance measure between the trajectory of objects in the two rollouts, or learned directly from pixels using a contrastive formulation. Empirically, we find that this approach leads to substantial performance improvements on the PHYRE benchmark for physical reasoning (Bakhtin et al., 2019), establishing a new state-of-the-art.
We propose UniT, a Unified Transformer model to simultaneously learn the most prominent tasks across different domains, ranging from object detection to natural language understanding and multimodal reasoning. Based on the transformer encoder-decoder architecture, our UniT model encodes each input modality with an encoder and makes predictions on each task with a shared decoder over the encoded input representations, followed by task-specific output heads. The entire model is jointly trained end-to-end with losses from each task. Compared to previous efforts on multi-task learning with transformers, we share the same model parameters across all tasks instead of separately fine-tuning task-specific models and handle a much higher variety of tasks across different domains. In our experiments, we learn 7 tasks jointly over 8 datasets, achieving strong performance on each task with significantly fewer parameters. Our code is available in MMF at https://mmf.sh.
Training scene graph classification models requires a large amount of annotated image data. Meanwhile, scene graphs represent relational knowledge that can be modeled with symbolic data from texts or knowledge graphs. While image annotation demands extensive labor, collecting textual descriptions of natural scenes requires less effort. In this work, we investigate whether textual scene descriptions can substitute for annotated image data. To this end, we employ a scene graph classification framework that is trained not only from annotated images but also from symbolic data. In our architecture, the symbolic entities are first mapped to their correspondent image-grounded representations and then fed into the relational reasoning pipeline. Even though a structured form of knowledge, such as the form in knowledge graphs, is not always available, we can generate it from unstructured texts using a transformer-based language model. We show that by fine-tuning the classification pipeline with the extracted knowledge from texts, we can achieve ~8x more accurate results in scene graph classification, ~3x in object classification, and ~1.5x in predicate classification, compared to the supervised baselines with only 1% of the annotated images.
Vision-and-Language Pre-training (VLP) has improved performance on various joint vision-and-language downstream tasks. Current approaches to VLP heavily rely on image feature extraction processes, most of which involve region supervision (e.g., object detection) and the convolutional architecture (e.g., ResNet). Although disregarded in the literature, we find it problematic in terms of both (1) efficiency/speed, that simply extracting input features requires much more computation than the multimodal interaction steps; and (2) expressive power, as it is upper bounded to the expressive power of the visual embedder and its predefined visual vocabulary. In this paper, we present a minimal VLP model, Vision-and-Language Transformer (ViLT), monolithic in the sense that the processing of visual inputs is drastically simplified to just the same convolution-free manner that we process textual inputs. We show that ViLT is up to tens of times faster than previous VLP models, yet with competitive or better downstream task performance. Our code and pre-trained weights are available at https://github.com/dandelin/vilt.