We present ViLBERT (short for Vision-and-Language BERT), a model for learning task-agnostic joint representations of image content and natural language. We extend the popular BERT architecture to a multi-modal two-stream model, pro-cessing both visual and textual inputs in separate streams that interact through co-attentional transformer layers. We pretrain our model through two proxy tasks on the large, automatically collected Conceptual Captions dataset and then transfer it to multiple established vision-and-language tasks -- visual question answering, visual commonsense reasoning, referring expressions, and caption-based image retrieval -- by making only minor additions to the base architecture. We observe significant improvements across tasks compared to existing task-specific models -- achieving state-of-the-art on all four tasks. Our work represents a shift away from learning groundings between vision and language only as part of task training and towards treating visual grounding as a pretrainable and transferable capability.
One of the primary challenges faced by deep learning is the degree to which current methods exploit superficial statistics and dataset bias, rather than learning to generalise over the specific representations they have experienced. This is a critical concern because generalisation enables robust reasoning over unseen data, whereas leveraging superficial statistics is fragile to even small changes in data distribution. To illuminate the issue and drive progress towards a solution, we propose a test that explicitly evaluates abstract reasoning over visual data. We introduce a large-scale benchmark of visual questions that involve operations fundamental to many high-level vision tasks, such as comparisons of counts and logical operations on complex visual properties. The benchmark directly measures a method's ability to infer high-level relationships and to generalise them over image-based concepts. It includes multiple training/test splits that require controlled levels of generalization. We evaluate a range of deep learning architectures, and find that existing models, including those popular for vision-and-language tasks, are unable to solve seemingly-simple instances. Models using relational networks fare better but leave substantial room for improvement.
We present a simple method that achieves unexpectedly superior performance for Complex Reasoning involved Visual Question Answering. Our solution collects statistical features from high-frequency words of all the questions asked about an image and use them as accurate knowledge for answering further questions of the same image. We are fully aware that this setting is not ubiquitously applicable, and in a more common setting one should assume the questions are asked separately and they cannot be gathered to obtain a knowledge base. Nonetheless, we use this method as an evidence to demonstrate our observation that the bottleneck effect is more severe on the feature extraction part than it is on the knowledge reasoning part. We show significant gaps when using the same reasoning model with 1) ground-truth features; 2) statistical features; 3) detected features from completely learned detectors, and analyze what these gaps mean to researches on visual reasoning topics. Our model with the statistical features achieves the 2nd place in the GQA Challenge 2019.
We introduce the Neural State Machine, seeking to bridge the gap between the neural and symbolic views of AI and integrate their complementary strengths for the task of visual reasoning. Given an image, we first predict a probabilistic graph that represents its underlying semantics and serves as a structured world model. Then, we perform sequential reasoning over the graph, iteratively traversing its nodes to answer a given question or draw a new inference. In contrast to most neural architectures that are designed to closely interact with the raw sensory data, our model operates instead in an abstract latent space, by transforming both the visual and linguistic modalities into semantic concept-based representations, thereby achieving enhanced transparency and modularity. We evaluate our model on VQA-CP and GQA, two recent VQA datasets that involve compositionality, multi-step inference and diverse reasoning skills, achieving state-of-the-art results in both cases. We provide further experiments that illustrate the model's strong generalization capacity across multiple dimensions, including novel compositions of concepts, changes in the answer distribution, and unseen linguistic structures, demonstrating the qualities and efficacy of our approach.
Grounding natural language in images, such as localizing "the black dog on the left of the tree", is one of the core problems in artificial intelligence, as it needs to comprehend the fine-grained and compositional language space. However, existing solutions rely on the association between the holistic language features and visual features, while neglect the nature of compositional reasoning implied in the language. In this paper, we propose a natural language grounding model that can automatically compose a binary tree structure for parsing the language and then perform visual reasoning along the tree in a bottom-up fashion. We call our model RVG-TREE: Recursive Grounding Tree, which is inspired by the intuition that any language expression can be recursively decomposed into two constituent parts, and the grounding confidence score can be recursively accumulated by calculating their grounding scores returned by sub-trees. RVG-TREE can be trained end-to-end by using the Straight-Through Gumbel-Softmax estimator that allows the gradients from the continuous score functions passing through the discrete tree construction. Experiments on several benchmarks show that our model achieves the state-of-the-art performance with more explainable reasoning.
A disentangled representation encodes information about the salient factors of variation in the data independently. Although it is often argued that this representational format is useful in learning to solve many real-world down-stream tasks, there is little empirical evidence that supports this claim. In this paper, we conduct a large-scale study that investigates whether disentangled representations are more suitable for abstract reasoning tasks. Using two new tasks similar to Raven's Progressive Matrices, we evaluate the usefulness of the representations learned by 360 state-of-the-art unsupervised disentanglement models. Based on these representations, we train 3600 abstract reasoning models and observe that disentangled representations do in fact lead to better down-stream performance. In particular, they enable quicker learning using fewer samples.
Without relevant human priors, neural networks may learn uninterpretable features. We propose Dynamics of Attention for Focus Transition (DAFT) as a human prior for machine reasoning. DAFT is a novel method that regularizes attention-based reasoning by modelling it as a continuous dynamical system using neural ordinary differential equations. As a proof of concept, we augment a state-of-the-art visual reasoning model with DAFT. Our experiments reveal that applying DAFT yields similar performance to the original model while using fewer reasoning steps, showing that it implicitly learns to skip unnecessary steps. We also propose a new metric, Total Length of Transition (TLT), which represents the effective reasoning step size by quantifying how much a given model's focus drifts while reasoning about a question. We show that adding DAFT results in lower TLT, demonstrating that our method indeed obeys the human prior towards shorter reasoning paths in addition to producing more interpretable attention maps. Our code is available at https://github.com/kakao/DAFT.
This paper presents a strong baseline for real-world visual reasoning (GQA), which achieves 60.93% in GQA 2019 challenge and won the sixth place. GQA is a large dataset with 22M questions involving spatial understanding and multi-step inference. To help further research in this area, we identified three crucial parts that improve the performance, namely: multi-source features, fine-grained encoder, and score-weighted ensemble. We provide a series of analysis on their impact on performance.
We do not speak word by word from scratch; our brain quickly structures a pattern like \textsc{sth do sth at someplace} and then fill in the detailed descriptions. To render existing encoder-decoder image captioners such human-like reasoning, we propose a novel framework: learning to Collocate Neural Modules (CNM), to generate the `inner pattern' connecting visual encoder and language decoder. Unlike the widely-used neural module networks in visual Q\&A, where the language (ie, question) is fully observable, CNM for captioning is more challenging as the language is being generated and thus is partially observable. To this end, we make the following technical contributions for CNM training: 1) compact module design --- one for function words and three for visual content words (eg, noun, adjective, and verb), 2) soft module fusion and multi-step module execution, robustifying the visual reasoning in partial observation, 3) a linguistic loss for module controller being faithful to part-of-speech collocations (eg, adjective is before noun). Extensive experiments on the challenging MS-COCO image captioning benchmark validate the effectiveness of our CNM image captioner. In particular, CNM achieves a new state-of-the-art 127.9 CIDEr-D on Karpathy split and a single-model 126.0 c40 on the official server. CNM is also robust to few training samples, eg, by training only one sentence per image, CNM can halve the performance loss compared to a strong baseline.
This paper studies the task of Visual Question Answering (VQA), which is topical in Multimedia community recently. Particularly, we explore two critical research problems existed in VQA: (1) efficiently fusing the visual and textual modalities; (2) enabling the visual reasoning ability of VQA models in answering complex questions. To address these challenging problems, a novel Question Guided Modular Routing Networks (QGMRN) has been proposed in this paper. Particularly, The QGMRN is composed of visual, textual and routing network. The visual and textual network serve as the backbones for the generic feature extractors of visual and textual modalities. QGMRN can fuse the visual and textual modalities at multiple semantic levels. Typically, the visual reasoning is facilitated by the routing network in a discrete and stochastic way by using Gumbel-Softmax trick for module selection. When the input reaches a certain modular layer, routing network newly proposed in this paper, dynamically selects a portion of modules from that layer to process the input depending on the question features generated by the textual network. It can also learn to reason by routing between the generic modules without additional supervision information or expert knowledge. Benefiting from the dynamic routing mechanism, QGMRN can outperform the previous classical VQA methods by a large margin and achieve the competitive results against the state-of-the-art methods. Furthermore, attention mechanism is integrated into our QGMRN model and thus can further boost the model performance. Empirically, extensive experiments on the CLEVR and CLEVR-Humans datasets validate the effectiveness of our proposed model, and the state-of-the-art performance has been achieved.
Dramatic progress has been witnessed in basic vision tasks involving low-level perception, such as object recognition, detection, and tracking. Unfortunately, there is still an enormous performance gap between artificial vision systems and human intelligence in terms of higher-level vision problems, especially ones involving reasoning. Earlier attempts in equipping machines with high-level reasoning have hovered around Visual Question Answering (VQA), one typical task associating vision and language understanding. In this work, we propose a new dataset, built in the context of Raven's Progressive Matrices (RPM) and aimed at lifting machine intelligence by associating vision with structural, relational, and analogical reasoning in a hierarchical representation. Unlike previous works in measuring abstract reasoning using RPM, we establish a semantic link between vision and reasoning by providing structure representation. This addition enables a new type of abstract reasoning by jointly operating on the structure representation. Machine reasoning ability using modern computer vision is evaluated in this newly proposed dataset. Additionally, we also provide human performance as a reference. Finally, we show consistent improvement across all models by incorporating a simple neural module that combines visual understanding and structure reasoning.
We introduce the new task of Acoustic Question Answering (AQA) to promote research in acoustic reasoning. The AQA task consists of analyzing an acoustic scene composed by a combination of elementary sounds and answering questions that relate the position and properties of these sounds. The kind of relational questions asked, require that the models perform non-trivial reasoning in order to answer correctly. Although similar problems have been extensively studied in the domain of visual reasoning, we are not aware of any previous studies addressing the problem in the acoustic domain. We propose a method for generating the acoustic scenes from elementary sounds and a number of relevant questions for each scene using templates. We also present preliminary results obtained with two models (FiLM and MAC) that have been shown to work for visual reasoning.
Reasoning about complex visual scenes involves perception of entities and their relations. Scene graphs provide a natural representation for reasoning tasks, by assigning labels to both entities (nodes) and relations (edges). Unfortunately, reasoning systems based on SGs are typically trained in a two-step procedure: First, training a model to predict SGs from images; Then, a separate model is created to reason based on predicted SGs. In many domains, it is preferable to train systems jointly in an end-to-end manner, but SGs are not commonly used as intermediate components in visual reasoning systems because being discrete and sparse, scene-graph representations are non-differentiable and difficult to optimize. Here we propose Differentiable Scene Graphs (DSGs), an image representation that is amenable to differentiable end-to-end optimization, and requires supervision only from the downstream tasks. DSGs provide a dense representation for all regions and pairs of regions, and do not spend modelling capacity on areas of the images that do not contain objects or relations of interest. We evaluate our model on the challenging task of identifying referring relationships (RR) in three benchmark datasets, Visual Genome, VRD and CLEVR. We describe a multi-task objective, and train in an end-to-end manner supervised by the downstream RR task. Using DSGs as an intermediate representation leads to new state-of-the-art performance.
We introduce GQA, a new dataset for real-world visual reasoning and compositional question answering, seeking to address key shortcomings of previous VQA datasets. We have developed a strong and robust question engine that leverages scene graph structures to create 22M diverse reasoning questions, all come with functional programs that represent their semantics. We use the programs to gain tight control over the answer distribution and present a new tunable smoothing technique to mitigate question biases. Accompanying the dataset is a suite of new metrics that evaluate essential qualities such as consistency, grounding and plausibility. An extensive analysis is performed for baselines as well as state-of-the-art models, providing fine-grained results for different question types and topologies. Whereas a blind LSTM obtains mere 42.1%, and strong VQA models achieve 54.1%, human performance tops at 89.3%, offering ample opportunity for new research to explore. We strongly hope GQA will provide an enabling resource for the next generation of models with enhanced robustness, improved consistency, and deeper semantic understanding for images and language.
We study the multi-round response generation in visual dialog, where a response is generated according to a visually grounded conversational history. Given a triplet: an image, Q&A history, and current question, all the prevailing methods follow a codec (i.e., encoder-decoder) fashion in a supervised learning paradigm: a multimodal encoder encodes the triplet into a feature vector, which is then fed into the decoder for the current answer generation, supervised by the ground-truth. However, this conventional supervised learning does NOT take into account the impact of imperfect history, violating the conversational nature of visual dialog and thus making the codec more inclined to learn history bias but not contextual reasoning. To this end, inspired by the actor-critic policy gradient in reinforcement learning, we propose a novel training paradigm called History Advantage Sequence Training (HAST). Specifically, we intentionally impose wrong answers in the history, obtaining an adverse critic, and see how the historic error impacts the codec's future behavior by History Advantage-a quantity obtained by subtracting the adverse critic from the gold reward of ground-truth history. Moreover, to make the codec more sensitive to the history, we propose a novel attention network called History-Aware Co-Attention Network (HACAN) which can be effectively trained by using HAST. Experimental results on three benchmarks: VisDial v0.9&v1.0 and GuessWhat?!, show that the proposed HAST strategy consistently outperforms the state-of-the-art supervised counterparts.
Learning to solve diagrammatic reasoning (DR) can be a challenging but interesting problem to the computer vision research community. It is believed that next generation pattern recognition applications should be able to simulate human brain to understand and analyze reasoning of images. However, due to the lack of benchmarks of diagrammatic reasoning, the present research primarily focuses on visual reasoning that can be applied to real-world objects. In this paper, we present a diagrammatic reasoning dataset that provides a large variety of DR problems. In addition, we also propose a Knowledge-based Long Short Term Memory (KLSTM) to solve diagrammatic reasoning problems. Our proposed analysis is arguably the first work in this research area. Several state-of-the-art learning frameworks have been used to compare with the proposed KLSTM framework in the present context. Preliminary results indicate that the domain is highly related to computer vision and pattern recognition research with several challenging avenues.
Discovering and exploiting the causality in deep neural networks (DNNs) are crucial challenges for understanding and reasoning causal effects (CE) on an explainable visual model. "Intervention" has been widely used for recognizing a causal relation ontologically. In this paper, we propose a causal inference framework for visual reasoning via do-calculus. To study the intervention effects on pixel-level features for causal reasoning, we introduce pixel-wise masking and adversarial perturbation. In our framework, CE is calculated using features in a latent space and perturbed prediction from a DNN-based model. We further provide the first look into the characteristics of discovered CE of adversarially perturbed images generated by gradient-based methods \footnote{~~https://github.com/jjaacckkyy63/Causal-Intervention-AE-wAdvImg}. Experimental results show that CE is a competitive and robust index for understanding DNNs when compared with conventional methods such as class-activation mappings (CAMs) on the Chest X-Ray-14 dataset for human-interpretable feature(s) (e.g., symptom) reasoning. Moreover, CE holds promises for detecting adversarial examples as it possesses distinct characteristics in the presence of adversarial perturbations.
Existing visual reasoning datasets such as Visual Question Answering (VQA), often suffer from biases conditioned on the question, image or answer distributions. The recently proposed CLEVR dataset addresses these limitations and requires fine-grained reasoning but the dataset is synthetic and consists of similar objects and sentence structures across the dataset. In this paper, we introduce a new inference task, Visual Entailment (VE) - consisting of image-sentence pairs whereby a premise is defined by an image, rather than a natural language sentence as in traditional Textual Entailment tasks. The goal of a trained VE model is to predict whether the image semantically entails the text. To realize this task, we build a dataset SNLI-VE based on the Stanford Natural Language Inference corpus and Flickr30k dataset. We evaluate various existing VQA baselines and build a model called Explainable Visual Entailment (EVE) system to address the VE task. EVE achieves up to 71% accuracy and outperforms several other state-of-the-art VQA based models. Finally, we demonstrate the explainability of EVE through cross-modal attention visualizations. The SNLI-VE dataset is publicly available at https://github.com/ necla-ml/SNLI-VE.
Deep Recurrent Neural Network (RNN) has gained popularity in many sequence classification tasks. Beyond predicting a correct class for each data instance, data scientists also want to understand what differentiating factors in the data have contributed to the classification during the learning process. We present a visual analytics approach to facilitate this task by revealing the RNN attention for all data instances, their temporal positions in the sequences, and the attribution of variables at each value level. We demonstrate with real-world datasets that our approach can help data scientists to understand such dynamics in deep RNNs from the training results, hence guiding their modeling process.
Referring object detection and referring image segmentation are important tasks that require joint understanding of visual information and natural language. Yet there has been evidence that current benchmark datasets suffer from bias, and current state-of-the-art models cannot be easily evaluated on their intermediate reasoning process. To address these issues and complement similar efforts in visual question answering, we build CLEVR-Ref+, a synthetic diagnostic dataset for referring expression comprehension. The precise locations and attributes of the objects are readily available, and the referring expressions are automatically associated with functional programs. The synthetic nature allows control over dataset bias (through sampling strategy), and the modular programs enable intermediate reasoning ground truth without human annotators. In addition to evaluating several state-of-the-art models on CLEVR-Ref+, we also propose IEP-Ref, a module network approach that significantly outperforms other models on our dataset. In particular, we present two interesting and important findings using IEP-Ref: (1) the module trained to transform feature maps into segmentation masks can be attached to any intermediate module to reveal the entire reasoning process step-by-step; (2) even if all training data has at least one object referred, IEP-Ref can correctly predict no-foreground when presented with false-premise referring expressions. To the best of our knowledge, this is the first direct and quantitative proof that neural modules behave in the way they are intended.