The open-ended question answering task of Text-VQA requires reading and reasoning about local, often previously unseen, scene-text content of an image to generate answers. In this work, we propose the generalized use of external knowledge to augment our understanding of the said scene-text. We design a framework to extract, filter, and encode knowledge atop a standard multimodal transformer for vision language understanding tasks. Through empirical evidence, we demonstrate how knowledge can highlight instance-only cues and thus help deal with training data bias, improve answer entity type correctness, and detect multiword named entities. We generate results comparable to the state-of-the-art on two publicly available datasets, under the constraints of similar upstream OCR systems and training data.
Current language models can generate high-quality text. Are they simply copying text they have seen before, or have they learned generalizable linguistic abstractions? To tease apart these possibilities, we introduce RAVEN, a suite of analyses for assessing the novelty of generated text, focusing on sequential structure (n-grams) and syntactic structure. We apply these analyses to four neural language models (an LSTM, a Transformer, Transformer-XL, and GPT-2). For local structure - e.g., individual dependencies - model-generated text is substantially less novel than our baseline of human-generated text from each model's test set. For larger-scale structure - e.g., overall sentence structure - model-generated text is as novel or even more novel than the human-generated baseline, but models still sometimes copy substantially, in some cases duplicating passages over 1,000 words long from the training set. We also perform extensive manual analysis showing that GPT-2's novel text is usually well-formed morphologically and syntactically but has reasonably frequent semantic issues (e.g., being self-contradictory).
Random masks define surprisingly effective sparse neural network models, as has been shown empirically. The resulting Erd\"os-R\'enyi (ER) random graphs can often compete with dense architectures and state-of-the-art lottery ticket pruning algorithms struggle to outperform them, even though the random baselines do not rely on computationally expensive pruning-training iterations but can be drawn initially without significant computational overhead. We offer a theoretical explanation of how such ER masks can approximate arbitrary target networks if they are wider by a logarithmic factor in the inverse sparsity $1 / \log(1/\text{sparsity})$. While we are the first to show theoretically and experimentally that random ER source networks contain strong lottery tickets, we also prove the existence of weak lottery tickets that require a lower degree of overparametrization than strong lottery tickets. These unusual results are based on the observation that ER masks are well trainable in practice, which we verify in experiments with varied choices of random masks. Some of these data-free choices outperform previously proposed random approaches on standard image classification benchmark datasets.
Long text understanding is important yet challenging in natural language processing. A long article or essay usually contains many redundant words that are not pertinent to its gist and sometimes can be regarded as noise. In this paper, we consider the problem of how to disentangle the gist-relevant and irrelevant information for long text understanding. With distillation mechanism, we transfer the knowledge about how to focus the salient parts from the abstractive summarization model and further integrate the distilled model, named \emph{Gist Detector}, into existing models as a supplementary component to augment the long text understanding. Experiments on document classification, distantly supervised open-domain question answering (DS-QA) and non-parallel text style transfer show that our method can significantly improve the performance of the baseline models, and achieves state-of-the-art overall results for document classification.
The attention-based encoder-decoder framework is becoming popular in scene text recognition, largely due to its superiority in integrating recognition clues from both visual and semantic domains. However, recent studies show the two clues might be misaligned in the difficult text (e.g., with rare text shapes) and introduce constraints such as character position to alleviate the problem. Despite certain success, a content-free positional embedding hardly associates with meaningful local image regions stably. In this paper, we propose a novel module called Multi-Domain Character Distance Perception (MDCDP) to establish a visual and semantic related position encoding. MDCDP uses positional embedding to query both visual and semantic features following the attention mechanism. It naturally encodes the positional clue, which describes both visual and semantic distances among characters. We develop a novel architecture named CDistNet that stacks MDCDP several times to guide precise distance modeling. Thus, the visual-semantic alignment is well built even various difficulties presented. We apply CDistNet to two augmented datasets and six public benchmarks. The experiments demonstrate that CDistNet achieves state-of-the-art recognition accuracy. While the visualization also shows that CDistNet achieves proper attention localization in both visual and semantic domains. We will release our code upon acceptance.
Potential societal and environmental effects such as the rapidly increasing resource use and the associated environmental impact, reproducibility issues, and exclusivity, the privatization of ML research leading to a public research brain-drain, a narrowing of the research effort caused by a focus on deep learning, and the introduction of biases through a lack of sociodemographic diversity in data and personnel caused by recent developments in machine learning are a current topic of discussion and scientific publications. However, these discussions and publications focus mainly on computer science-adjacent fields, including computer vision and natural language processing or basic ML research. Using bibliometric analysis of the complete and full-text analysis of the open-access literature, we show that the same observations can be made for applied machine learning in chemistry and biology. These developments can potentially affect basic and applied research, such as drug discovery and development, beyond the known issue of biased data sets.
The rapid development of Artificial Intelligence (AI) requires developers and designers of AI systems to focus on the collaboration between humans and machines. AI explanations of system behavior and reasoning are vital for effective collaboration by fostering appropriate trust, ensuring understanding, and addressing issues of fairness and bias. However, various contextual and subjective factors can influence an AI system explanation's effectiveness. This work draws inspiration from findings in cognitive psychology to understand how effective explanations can be designed. We identify four components to which explanation designers can pay special attention: perception, semantics, intent, and user & context. We illustrate the use of these four explanation components with an example of estimating food calories by combining text with visuals, probabilities with exemplars, and intent communication with both user and context in mind. We propose that the significant challenge for effective AI explanations is an additional step between explanation generation using algorithms not producing interpretable explanations and explanation communication. We believe this extra step will benefit from carefully considering the four explanation components outlined in our work, which can positively affect the explanation's effectiveness.
We examine the inducement of rare but severe errors in English-Chinese and Chinese-English in-domain neural machine translation by minimal deletion of the source text with character-based models. By deleting a single character, we can induce severe translation errors. We categorize these errors and compare the results of deleting single characters and single words. We also examine the effect of training data size on the number and types of pathological cases induced by these minimal perturbations, finding significant variation. We find that deleting a word hurts overall translation score more than deleting a character, but certain errors are more likely to occur when deleting characters, with language direction also influencing the effect.
Automated monitoring of dark web (DW) platforms on a large scale is the first step toward developing proactive Cyber Threat Intelligence (CTI). While there are efficient methods for collecting data from the surface web, large-scale dark web data collection is often hindered by anti-crawling measures. In particular, text-based CAPTCHA serves as the most prevalent and prohibiting type of these measures in the dark web. Text-based CAPTCHA identifies and blocks automated crawlers by forcing the user to enter a combination of hard-to-recognize alphanumeric characters. In the dark web, CAPTCHA images are meticulously designed with additional background noise and variable character length to prevent automated CAPTCHA breaking. Existing automated CAPTCHA breaking methods have difficulties in overcoming these dark web challenges. As such, solving dark web text-based CAPTCHA has been relying heavily on human involvement, which is labor-intensive and time-consuming. In this study, we propose a novel framework for automated breaking of dark web CAPTCHA to facilitate dark web data collection. This framework encompasses a novel generative method to recognize dark web text-based CAPTCHA with noisy background and variable character length. To eliminate the need for human involvement, the proposed framework utilizes Generative Adversarial Network (GAN) to counteract dark web background noise and leverages an enhanced character segmentation algorithm to handle CAPTCHA images with variable character length. Our proposed framework, DW-GAN, was systematically evaluated on multiple dark web CAPTCHA testbeds. DW-GAN significantly outperformed the state-of-the-art benchmark methods on all datasets, achieving over 94.4% success rate on a carefully collected real-world dark web dataset...
Recent advances in pre-training vision-language models like CLIP have shown great potential in learning transferable visual representations. Nonetheless, for downstream inference, CLIP-like models suffer from either 1) degraded accuracy and robustness in the case of inaccurate text descriptions during retrieval-based inference (the challenge for zero-shot protocol); or 2) breaking the well-established vision-language alignment (the challenge for linear probing). To address them, we propose Decomposed Feature Prompting (DeFo). DeFo leverages a flexible number of learnable embeddings as textual input while maintaining the vision-language dual-model architecture, which enables the model to learn decomposed visual features with the help of feature-level textual prompts. We further use an additional linear layer to perform classification, allowing a scalable size of language inputs. Our empirical study shows DeFo's significance in improving the vision-language models. For example, DeFo obtains 73.2% test accuracy on ImageNet with a ResNet-50 backbone without tuning any pretrained weights of both the vision and language encoder, outperforming zero-shot CLIP by a large margin of 15.0%, and outperforming state-of-the-art vision-language prompt tuning method by 7.6%.