Transformer-based models have pushed state of the art in many areas of NLP, but our understanding of what is behind their success is still limited. This paper is the first survey of over 150 studies of the popular BERT model. We review the current state of knowledge about how BERT works, what kind of information it learns and how it is represented, common modifications to its training objectives and architecture, the overparameterization issue and approaches to compression. We then outline directions for future research.

Recent advances in deep learning research, such as transformers, have bolstered the ability for automated agents to generate creative texts similar to those that a human would write. By default, transformer decoders can only generate new text with respect to previously generated text. The output distribution of candidate tokens at any position is conditioned on previously selected tokens using a self-attention mechanism to emulate the property of autoregression. This is inherently limiting for tasks such as controllable story generation where it may be necessary to condition on future plot events when writing a story. In this work, we propose Future Sight, a method for finetuning a pretrained generative transformer on the task of future conditioning. Transformer decoders are typically pretrained on the task of completing a context, one token at a time, by means of self-attention. Future Sight additionally enables a decoder to attend to an encoded future plot event. This motivates the decoder to expand on the context in a way that logically concludes with the provided future. During inference, the future plot event can be written by a human author to steer the narrative being generated in a certain direction. We evaluate the efficacy of our approach on a story generation task with human evaluators.

Current large language models can perform reasonably well on complex tasks that require step-by-step reasoning with few-shot learning. Are these models applying reasoning skills they have learnt during pre-training and reason outside of their training context, or are they simply memorizing their training corpus at finer granularity and have learnt to better understand their context? To tease apart these possibilities, we introduce ALERT, a benchmark and suite of analyses for assessing language models' reasoning ability comparing pre-trained and finetuned models on complex tasks that require reasoning skills to solve. ALERT provides a test bed to asses any language model on fine-grained reasoning skills, which spans over 20 datasets and covers 10 different reasoning skills. We leverage ALERT to further investigate the role of finetuning. With extensive empirical analysis we find that language models learn more reasoning skills such as textual entailment, abductive reasoning, and analogical reasoning during finetuning stage compared to pretraining state. We also find that when language models are finetuned they tend to overfit to the prompt template, which hurts the robustness of models causing generalization problems.

Large language models show improved downstream task performance when prompted to generate step-by-step reasoning to justify their final answers. These reasoning steps greatly improve model interpretability and verification, but objectively studying their correctness (independent of the final answer) is difficult without reliable methods for automatic evaluation. We simply do not know how often the stated reasoning steps actually support the final end task predictions. In this work, we present ROSCOE, a suite of interpretable, unsupervised automatic scores that improve and extend previous text generation evaluation metrics. To evaluate ROSCOE against baseline metrics, we design a typology of reasoning errors and collect synthetic and human evaluation scores on commonly used reasoning datasets. In contrast with existing metrics, ROSCOE can measure semantic consistency, logicality, informativeness, fluency, and factuality - among other traits - by leveraging properties of step-by-step rationales. We empirically verify the strength of our metrics on five human annotated and six programmatically perturbed diagnostics datasets - covering a diverse set of tasks that require reasoning skills and show that ROSCOE can consistently outperform baseline metrics.

Nowadays, copy detection patterns (CDP) appear as a very promising anti-counterfeiting technology for physical object protection. However, the advent of deep learning as a powerful attacking tool has shown that the general authentication schemes are unable to compete and fail against such attacks. In this paper, we propose a new mathematical model of printing-imaging channel for the authentication of CDP together with a new detection scheme based on it. The results show that even deep learning created copy fakes unknown at the training stage can be reliably authenticated based on the proposed approach and using only digital references of CDP during authentication.

Physics-Informed Neural Networks (PINNs) are gaining popularity as a method for solving differential equations. While being more feasible in some contexts than the classical numerical techniques, PINNs still lack credibility. A remedy for that can be found in Uncertainty Quantification (UQ) which is just beginning to emerge in the context of PINNs. Assessing how well the trained PINN complies with imposed differential equation is the key to tackling uncertainty, yet there is lack of comprehensive methodology for this task. We propose a framework for UQ in Bayesian PINNs (B-PINNs) that incorporates the discrepancy between the B-PINN solution and the unknown true solution. We exploit recent results on error bounds for PINNs on linear dynamical systems and demonstrate the predictive uncertainty on a class of linear ODEs.

Artificial Intelligence (AI) systems have been increasingly used to make decision-making processes faster, more accurate, and more efficient. However, such systems are also at constant risk of being attacked. While the majority of attacks targeting AI-based applications aim to manipulate classifiers or training data and alter the output of an AI model, recently proposed Sponge Attacks against AI models aim to impede the classifier's execution by consuming substantial resources. In this work, we propose \textit{Dual Denial of Decision (DDoD) attacks against collaborative Human-AI teams}. We discuss how such attacks aim to deplete \textit{both computational and human} resources, and significantly impair decision-making capabilities. We describe DDoD on human and computational resources and present potential risk scenarios in a series of exemplary domains.

The sheer volume of online user-generated content has rendered content moderation technologies essential in order to protect digital platform audiences from content that may cause anxiety, worry, or concern. Despite the efforts towards developing automated solutions to tackle this problem, creating accurate models remains challenging due to the lack of adequate task-specific training data. The fact that manually annotating such data is a highly demanding procedure that could severely affect the annotators' emotional well-being is directly related to the latter limitation. In this paper, we propose the CM-Refinery framework that leverages large-scale multimedia datasets to automatically extend initial training datasets with hard examples that can refine content moderation models, while significantly reducing the involvement of human annotators. We apply our method on two model adaptation strategies designed with respect to the different challenges observed while collecting data, i.e. lack of (i) task-specific negative data or (ii) both positive and negative data. Additionally, we introduce a diversity criterion applied to the data collection process that further enhances the generalization performance of the refined models. The proposed method is evaluated on the Not Safe for Work (NSFW) and disturbing content detection tasks on benchmark datasets achieving 1.32% and 1.94% accuracy improvements compared to the state of the art, respectively. Finally, it significantly reduces human involvement, as 92.54% of data are automatically annotated in case of disturbing content while no human intervention is required for the NSFW task.

In this paper we look into the conjecture of Entezari et al. (2021) which states that if the permutation invariance of neural networks is taken into account, then there is likely no loss barrier to the linear interpolation between SGD solutions. First, we observe that neuron alignment methods alone are insufficient to establish low-barrier linear connectivity between SGD solutions due to a phenomenon we call variance collapse: interpolated deep networks suffer a collapse in the variance of their activations, causing poor performance. Next, we propose REPAIR (REnormalizing Permuted Activations for Interpolation Repair) which mitigates variance collapse by rescaling the preactivations of such interpolated networks. We explore the interaction between our method and the choice of normalization layer, network width, and depth, and demonstrate that using REPAIR on top of neuron alignment methods leads to 60%-100% relative barrier reduction across a wide variety of architecture families and tasks. In particular, we report a 74% barrier reduction for ResNet50 on ImageNet and 90% barrier reduction for ResNet18 on CIFAR10.

Spatiotemporal data is readily available due to emerging sensor and data acquisition technologies that track the positions of moving objects of interest. Spatiotemporal clustering addresses the need to efficiently discover patterns and trends in moving object behavior without human supervision. One application of interest is the discovery of moving clusters, where clusters have a static identity, but their location and content can change over time. We propose a two phase spatiotemporal clustering method called spatiotemporal k-means (STKM) that is able to analyze the multi-scale relationships within spatiotemporal data. Phase 1 of STKM frames the moving cluster problem as the minimization of an objective function unified over space and time. It outputs the short-term associations between objects and is uniquely able to track dynamic cluster centers with minimal parameter tuning and without post-processing. Phase 2 outputs the long-term associations and can be applied to any method that provides a cluster label for each object at every point in time. We evaluate STKM against baseline methods on a recently developed benchmark dataset and show that STKM outperforms existing methods, particularly in the low-data domain, with significant performance improvements demonstrated for common evaluation metrics on the moving cluster problem.