The factual knowledge acquired during pretraining and stored in the parameters of Language Models (LM) can be useful in downstream tasks (e.g., question answering or textual inference). However, some facts can be incorrectly induced or become obsolete over time. We present KnowledgeEditor, a method that can be used to edit this knowledge and, thus, fix 'bugs' or unexpected predictions without the need for expensive re-training or fine-tuning. Besides being computationally efficient, KnowledgeEditor does not require any modifications in LM pre-training (e.g., the use of meta-learning). In our approach, we train a hyper-network with constrained optimization to modify a fact without affecting the rest of the knowledge; the trained hyper-network is then used to predict the weight update at test time. We show KnowledgeEditor's efficacy with two popular architectures and knowledge-intensive tasks: i) a BERT model fine-tuned for fact-checking, and ii) a sequence-to-sequence BART model for question answering. With our method, changing a prediction on the specific wording of a query tends to result in a consistent change in predictions also for its paraphrases. We show that this can be further encouraged by exploiting (e.g., automatically-generated) paraphrases during training. Interestingly, our hyper-network can be regarded as a 'probe' revealing which components of a model need to be changed to manipulate factual knowledge; our analysis shows that the updates tend to be concentrated on a small subset of components. Code at https://github.com/nicola-decao/KnowledgeEditor

Pretrained Masked Language Models (MLMs) have revolutionised NLP in recent years. However, previous work has indicated that off-the-shelf MLMs are not effective as universal lexical or sentence encoders without further task-specific fine-tuning on NLI, sentence similarity, or paraphrasing tasks using annotated task data. In this work, we demonstrate that it is possible to turn MLMs into effective universal lexical and sentence encoders even without any additional data and without any supervision. We propose an extremely simple, fast and effective contrastive learning technique, termed Mirror-BERT, which converts MLMs (e.g., BERT and RoBERTa) into such encoders in less than a minute without any additional external knowledge. Mirror-BERT relies on fully identical or slightly modified string pairs as positive (i.e., synonymous) fine-tuning examples, and aims to maximise their similarity during identity fine-tuning. We report huge gains over off-the-shelf MLMs with Mirror-BERT in both lexical-level and sentence-level tasks, across different domains and different languages. Notably, in the standard sentence semantic similarity (STS) tasks, our self-supervised Mirror-BERT model even matches the performance of the task-tuned Sentence-BERT models from prior work. Finally, we delve deeper into the inner workings of MLMs, and suggest some evidence on why this simple approach can yield effective univeral lexical and sentence encoders.

In Machine Learning, the $\mathsf{SHAP}$-score is a version of the Shapley value that is used to explain the result of a learned model on a specific entity by assigning a score to every feature. While in general computing Shapley values is an intractable problem, we prove a strong positive result stating that the $\mathsf{SHAP}$-score can be computed in polynomial time over deterministic and decomposable Boolean circuits. Such circuits are studied in the field of Knowledge Compilation and generalize a wide range of Boolean circuits and binary decision diagrams classes, including binary decision trees and Ordered Binary Decision Diagrams (OBDDs). We also establish the computational limits of the SHAP-score by observing that computing it over a class of Boolean models is always polynomially as hard as the model counting problem for that class. This implies that both determinism and decomposability are essential properties for the circuits that we consider. It also implies that computing $\mathsf{SHAP}$-scores is intractable as well over the class of propositional formulas in DNF. Based on this negative result, we look for the existence of fully-polynomial randomized approximation schemes (FPRAS) for computing $\mathsf{SHAP}$-scores over such class. In contrast to the model counting problem for DNF formulas, which admits an FPRAS, we prove that no such FPRAS exists for the computation of $\mathsf{SHAP}$-scores. Surprisingly, this negative result holds even for the class of monotone formulas in DNF. These techniques can be further extended to prove another strong negative result: Under widely believed complexity assumptions, there is no polynomial-time algorithm that checks, given a monotone DNF formula $\varphi$ and features $x,y$, whether the $\mathsf{SHAP}$-score of $x$ in $\varphi$ is smaller than the $\mathsf{SHAP}$-score of $y$ in $\varphi$.

Large pretrained language models like GPT-3 have acquired a surprising ability to perform zero-shot classification (ZSC). For example, to classify review sentiments, we can "prompt" the language model with the review and the question "Is the review positive?" as the context, and ask it to predict whether the next word is "Yes" or "No". However, these models are not specialized for answering these prompts. To address this weakness, we propose meta-tuning, which trains the model to specialize in answering prompts but still generalize to unseen tasks. To create the training data, we aggregated 43 existing datasets, annotated 441 label descriptions in total, and unified them into the above question answering (QA) format. After meta-tuning, our model outperforms a same-sized QA model for most labels on unseen tasks, and we forecast that the performance would improve for even larger models. Therefore, measuring ZSC performance on non-specialized language models might underestimate their true capability, and community-wide efforts on aggregating datasets and unifying their formats can help build models that understand prompts better.

It all started with an odd pile of shells: a pile that, upon closer inspection, fell apart like a flower losing its petals, introducing a burned-out nature documentarian named Craig Foster--and, in time, the world--to the octopus hiding cleverly inside. Known simply as "her," she would become the star of My Octopus Teacher, the Oscar-nominated Netflix documentary and surprise pandemic hit that told the story of Foster's unlikely relationship with that eight-armed mollusk. Released in September 2020, it arrived at the perfect moment. Audiences exhausted by lockdowns and unrelenting 2020-ness were primed for escape into the undersea fantasia of South Africa's kelp forests, where Foster met her. Best-selling books like The Soul of an Octopus and Other Minds: The Octopus, the Sea, and the Deep Origins of Consciousness had whetted public curiosity about these uncannily intelligent creatures with whom humans last shared a common ancestor 600 million years ago. Yet while most writing about octopuses emphasizes their ostensibly alien, unknowable nature,1 and serious, science-minded nature documentaries elevate concern about biodiversity over sentiment for a single animal, My Octopus Teacher defied convention. It embraced Foster's feelings for the octopus, which over the course of a year evolved from curiosity to care--even to love. And though her own feelings were left for viewers to interpret, the film's indelible impression was of nature populated by species who are not only beautiful and exquisitely evolved and ecologically important, but highly sentient, too. Nautilus talked to Foster about his octopus teacher and how getting to know her changed the way he thinks about nature. I write a lot about nature and biology and ecology, but in the last few years I've focused on the minds of animals and how we think about them.

Training with Reinforcement Learning requires a reward function that is used to guide the agent towards achieving its objective. However, designing smooth and well-behaved rewards is in general not trivial and requires significant human engineering efforts. Generating rewards in self-supervised way, by inspiring the agent with an intrinsic desire to learn and explore the environment, might induce more general behaviours. In this work, we propose a curiosity-based bonus as intrinsic reward for Reinforcement Learning, computed as the Bayesian surprise with respect to a latent state variable, learnt by reconstructing fixed random features. We extensively evaluate our model by measuring the agent's performance in terms of environment exploration, for continuous tasks, and looking at the game scores achieved, for video games. Our model is computationally cheap and empirically shows state-of-the-art performance on several problems. Furthermore, experimenting on an environment with stochastic actions, our approach emerged to be the most resilient to simple stochasticity. Further visualization is available on the project webpage.(https://lbsexploration.github.io/)

We consider the problem of jointly modeling and clustering populations of tensors by introducing a flexible high-dimensional tensor mixture model with heterogeneous covariances. The proposed mixture model exploits the intrinsic structures of tensor data, and is assumed to have means that are low-rank and internally sparse as well as heterogeneous covariances that are separable and conditionally sparse. We develop an efficient high-dimensional expectation-conditional-maximization (HECM) algorithm that breaks the challenging optimization in the M-step into several simpler conditional optimization problems, each of which is convex, admits regularization and has closed-form updating formulas. We show that the proposed HECM algorithm, with an appropriate initialization, converges geometrically to a neighborhood that is within statistical precision of the true parameter. Such a theoretical analysis is highly nontrivial due to the dual non-convexity arising from both the EM-type estimation and the non-convex objective function in the M-step. The efficacy of our proposed method is demonstrated through simulation studies and an application to an autism spectrum disorder study, where our analysis identifies important brain regions for diagnosis.

Recent work has shown that commonly available machine reading comprehension (MRC) datasets can be used to train high-performance neural information retrieval (IR) systems. However, the evaluation of neural IR has so far been limited to standard supervised learning settings, where they have outperformed traditional term matching baselines. We conduct in-domain and out-of-domain evaluations of neural IR, and seek to improve its robustness across different scenarios, including zero-shot settings. We show that synthetic training examples generated using a sequence-to-sequence generator can be effective towards this goal: in our experiments, pre-training with synthetic examples improves retrieval performance in both in-domain and out-of-domain evaluation on five different test sets.

It uses comparative empirical experiments (Charrette Test Method) to draw comparisons between the visibility of state-of-practice and blockchain-enabled payment systems in a commercial construction project. Comparisons were drawn across four levels of granularity. The findings are twofold: 1) blockchain improved information completeness and information accuracy respectively by an average 216% and 261% compared with the digital state-of-practice solution. The improvements were significantly more pronounced for inquiries that had higher product, trade, and temporal granularity; 2) blockchain-enabled solution was robust in the face of increased granularity, while the conventional solution experienced 50% and 66.7% decline respectively in completeness and accuracy of information. The paper concludes with a discussion of mechanisms contributing to visibility and technology adoption based on business objectives.

Machine learning has brought striking advances in multilingual natural language processing capabilities over the past year. For example, the latest techniques have improved the state-of-the-art performance on the XTREME multilingual benchmark by more than 13 points. While a sizeable gap to human-level performance remains, improvements have been easier to achieve in some tasks than in others. This paper analyzes the current state of cross-lingual transfer learning and summarizes some lessons learned. In order to catalyze meaningful progress, we extend XTREME to XTREME-R, which consists of an improved set of ten natural language understanding tasks, including challenging language-agnostic retrieval tasks, and covers 50 typologically diverse languages. In addition, we provide a massively multilingual diagnostic suite and fine-grained multi-dataset evaluation capabilities through an interactive public leaderboard to gain a better understanding of such models.