State-of-the-art image and text classification models, such as Convectional Neural Networks and Transformers, have long been able to classify their respective unimodal reasoning satisfactorily with accuracy close to or exceeding human accuracy. However, images embedded with text, such as hateful memes, are hard to classify using unimodal reasoning when difficult examples, such as benign confounders, are incorporated into the data set. We attempt to generate more labeled memes in addition to the Hateful Memes data set from Facebook AI, based on the framework of a winning team from the Hateful Meme Challenge. To increase the number of labeled memes, we explore semi-supervised learning using pseudo-labels for newly introduced, unlabeled memes gathered from the Memotion Dataset 7K. We find that the semi-supervised learning task on unlabeled data required human intervention and filtering and that adding a limited amount of new data yields no extra classification performance.
相關內容
Collecting large-scale datasets is crucial for training deep models, annotating the data, however, inevitably yields noisy labels, which poses challenges to deep learning algorithms. Previous efforts tend to mitigate this problem via identifying and removing noisy samples or correcting their labels according to the statistical properties (e.g., loss values) among training samples. In this paper, we aim to tackle this problem from a new perspective, delving into the deep feature maps, we empirically find that models trained with clean and mislabeled samples manifest distinguishable activation feature distributions. From this observation, a novel robust training approach termed adversarial noisy masking is proposed. The idea is to regularize deep features with a label quality guided masking scheme, which adaptively modulates the input data and label simultaneously, preventing the model to overfit noisy samples. Further, an auxiliary task is designed to reconstruct input data, it naturally provides noise-free self-supervised signals to reinforce the generalization ability of deep models. The proposed method is simple and flexible, it is tested on both synthetic and real-world noisy datasets, where significant improvements are achieved over previous state-of-the-art methods.
Although backdoor learning is an active research topic in the NLP domain, the literature lacks studies that systematically categorize and summarize backdoor attacks and defenses. To bridge the gap, we present a comprehensive and unifying study of backdoor learning for NLP by summarizing the literature in a systematic manner. We first present and motivate the importance of backdoor learning for building robust NLP systems. Next, we provide a thorough account of backdoor attack techniques, their applications, defenses against backdoor attacks, and various mitigation techniques to remove backdoor attacks. We then provide a detailed review and analysis of evaluation metrics, benchmark datasets, threat models, and challenges related to backdoor learning in NLP. Ultimately, our work aims to crystallize and contextualize the landscape of existing literature in backdoor learning for the text domain and motivate further research in the field. To this end, we identify troubling gaps in the literature and offer insights and ideas into open challenges and future research directions. Finally, we provide a GitHub repository with a list of backdoor learning papers that will be continuously updated at //github.com/marwanomar1/Backdoor-Learning-for-NLP.
Supervised machine learning utilizes large datasets, often with ground truth labels annotated by humans. While some data points are easy to classify, others are hard to classify, which reduces the inter-annotator agreement. This causes noise for the classifier and might affect the user's perception of the classifier's performance. In our research, we investigated whether the classification difficulty of a data point influences how strongly a prediction mistake reduces the "perceived accuracy". In an experimental online study, 225 participants interacted with three fictive classifiers with equal accuracy (73%). The classifiers made prediction mistakes on three different types of data points (easy, difficult, impossible). After the interaction, participants judged the classifier's accuracy. We found that not all prediction mistakes reduced the perceived accuracy equally. Furthermore, the perceived accuracy differed significantly from the calculated accuracy. To conclude, accuracy and related measures seem unsuitable to represent how users perceive the performance of classifiers.
Exploiting social media to spread hate has tremendously increased over the years. Lately, multi-modal hateful content such as memes has drawn relatively more traction than uni-modal content. Moreover, the availability of implicit content payloads makes them fairly challenging to be detected by existing hateful meme detection systems. In this paper, we present a use case study to analyze such systems' vulnerabilities against external adversarial attacks. We find that even very simple perturbations in uni-modal and multi-modal settings performed by humans with little knowledge about the model can make the existing detection models highly vulnerable. Empirically, we find a noticeable performance drop of as high as 10% in the macro-F1 score for certain attacks. As a remedy, we attempt to boost the model's robustness using contrastive learning as well as an adversarial training-based method - VILLA. Using an ensemble of the above two approaches, in two of our high resolution datasets, we are able to (re)gain back the performance to a large extent for certain attacks. We believe that ours is a first step toward addressing this crucial problem in an adversarial setting and would inspire more such investigations in the future.
This paper presents a data-driven framework to improve the trustworthiness of US tax preparation software systems. Given the legal implications of bugs in such software on its users, ensuring compliance and trustworthiness of tax preparation software is of paramount importance. The key barriers in developing debugging aids for tax preparation systems are the unavailability of explicit specifications and the difficulty of obtaining oracles. We posit that, since the US tax law adheres to the legal doctrine of precedent, the specifications about the outcome of tax preparation software for an individual taxpayer must be viewed in comparison with individuals that are deemed similar. Consequently, these specifications are naturally available as properties on the software requiring similar inputs provide similar outputs. Inspired by the metamorphic testing paradigm, we dub these relations metamorphic relations. In collaboration with legal and tax experts, we explicated metamorphic relations for a set of challenging properties from various US Internal Revenue Services (IRS) publications including Publication 596 (Earned Income Tax Credit), Schedule 8812 (Qualifying Children/Other Dependents), and Form 8863 (Education Credits). We focus on an open-source tax preparation software for our case study and develop a randomized test-case generation strategy to systematically validate the correctness of tax preparation software guided by metamorphic relations. We further aid this test-case generation by visually explaining the behavior of software on suspicious instances using easy to-interpret decision-tree models. Our tool uncovered several accountability bugs with varying severity ranging from non-robust behavior in corner-cases (unreliable behavior when tax returns are close to zero) to missing eligibility conditions in the updated versions of software.
An ongoing challenge in current natural language processing is how its major advancements tend to disproportionately favor resource-rich languages, leaving a significant number of under-resourced languages behind. Due to the lack of resources required to train and evaluate models, most modern language technologies are either nonexistent or unreliable to process endangered, local, and non-standardized languages. Optical character recognition (OCR) is often used to convert endangered language documents into machine-readable data. However, such OCR output is typically noisy, and most word alignment models are not built to work under such noisy conditions. In this work, we study the existing word-level alignment models under noisy settings and aim to make them more robust to noisy data. Our noise simulation and structural biasing method, tested on multiple language pairs, manages to reduce the alignment error rate on a state-of-the-art neural-based alignment model up to 59.6%.
Labeling data is one of the most costly processes in machine learning pipelines. Active learning is a standard approach to alleviating this problem. Pool-based active learning first builds a pool of unlabelled data and iteratively selects data to be labeled so that the total number of required labels is minimized, keeping the model performance high. Many effective criteria for choosing data from the pool have been proposed in the literature. However, how to build the pool is less explored. Specifically, most of the methods assume that a task-specific pool is given for free. In this paper, we advocate that such a task-specific pool is not always available and propose the use of a myriad of unlabelled data on the Web for the pool for which active learning is applied. As the pool is extremely large, it is likely that relevant data exist in the pool for many tasks, and we do not need to explicitly design and build the pool for each task. The challenge is that we cannot compute the acquisition scores of all data exhaustively due to the size of the pool. We propose an efficient method, Seafaring, to retrieve informative data in terms of active learning from the Web using a user-side information retrieval algorithm. In the experiments, we use the online Flickr environment as the pool for active learning. This pool contains more than ten billion images and is several orders of magnitude larger than the existing pools in the literature for active learning. We confirm that our method performs better than existing approaches of using a small unlabelled pool.
We explore combining batch order-fair atomic broadcast (of-ABC) and frequent batch auction (FBA) as a defense against general order manipulations in blockchain-based decentralized exchanges (DEX). To justify FBA, we compare the welfare loss of decentralized exchanges under two market designs: continuous limit order book (CLOB), where transactions are processed sequentially, and FBA, where transactions are arranged into batches and a uniform price double auction decides execution order. We model three types of players, common investors, privately informed traders, and arbitrageurs who can provide liquidity and front-run, along with a decentralized exchange. Assuming that the exchange is realized over an of-ABC protocol, we find that FBA can achieve better social welfare compared to CLOB when (1) public information affecting the fundamental value of an asset is revealed more frequently, or (2) the block generation interval is sufficiently large, or (3) the priority fees are small compared to the asset price changes, or (4) fewer privately informed parties exist. Intrinsic reasons are that first, blockchains already treat time as discrete and ensuring order fairness there is non-trivial, allowing even more room for latency arbitrage rents under CLOB; second, sufficiently large block creation interval allows for information dispersion; third, higher priority fees discourage front-running under CLOB; additionally, FBA prioritizes price in deciding execution order and fewer informed traders mean less adverse price impact.
The canonical approach to video-and-language learning (e.g., video question answering) dictates a neural model to learn from offline-extracted dense video features from vision models and text features from language models. These feature extractors are trained independently and usually on tasks different from the target domains, rendering these fixed features sub-optimal for downstream tasks. Moreover, due to the high computational overload of dense video features, it is often difficult (or infeasible) to plug feature extractors directly into existing approaches for easy finetuning. To provide a remedy to this dilemma, we propose a generic framework ClipBERT that enables affordable end-to-end learning for video-and-language tasks, by employing sparse sampling, where only a single or a few sparsely sampled short clips from a video are used at each training step. Experiments on text-to-video retrieval and video question answering on six datasets demonstrate that ClipBERT outperforms (or is on par with) existing methods that exploit full-length videos, suggesting that end-to-end learning with just a few sparsely sampled clips is often more accurate than using densely extracted offline features from full-length videos, proving the proverbial less-is-more principle. Videos in the datasets are from considerably different domains and lengths, ranging from 3-second generic domain GIF videos to 180-second YouTube human activity videos, showing the generalization ability of our approach. Comprehensive ablation studies and thorough analyses are provided to dissect what factors lead to this success. Our code is publicly available at //github.com/jayleicn/ClipBERT
The Q-learning algorithm is known to be affected by the maximization bias, i.e. the systematic overestimation of action values, an important issue that has recently received renewed attention. Double Q-learning has been proposed as an efficient algorithm to mitigate this bias. However, this comes at the price of an underestimation of action values, in addition to increased memory requirements and a slower convergence. In this paper, we introduce a new way to address the maximization bias in the form of a "self-correcting algorithm" for approximating the maximum of an expected value. Our method balances the overestimation of the single estimator used in conventional Q-learning and the underestimation of the double estimator used in Double Q-learning. Applying this strategy to Q-learning results in Self-correcting Q-learning. We show theoretically that this new algorithm enjoys the same convergence guarantees as Q-learning while being more accurate. Empirically, it performs better than Double Q-learning in domains with rewards of high variance, and it even attains faster convergence than Q-learning in domains with rewards of zero or low variance. These advantages transfer to a Deep Q Network implementation that we call Self-correcting DQN and which outperforms regular DQN and Double DQN on several tasks in the Atari 2600 domain.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191