Recent advances in text-to-image synthesis make it possible to visualize machine imaginations for a given context. On the other hand, when generating text, human writers are gifted at creative visualization, which enhances their writings by forming imaginations as blueprints before putting down the stories in words. Inspired by such a cognitive process, we ask the natural question of whether we can endow machines with the same ability to utilize visual information and construct a general picture of the context to guide text generation. In this work, we propose iNLG that uses machine-generated images to guide language models in open-ended text generation. The experiments and analyses demonstrate the effectiveness of iNLG on open-ended text generation tasks, including text completion, story generation, and concept-to-text generation in both few-shot and full-data scenarios. Both automatic metrics and human evaluations verify that the text snippets generated by our iNLG are coherent and informative while displaying minor degeneration.

Little attention is placed on analyzing nationality bias in language models, especially when nationality is highly used as a factor in increasing the performance of social NLP models. This paper examines how a text generation model, GPT-2, accentuates pre-existing societal biases about country-based demonyms. We generate stories using GPT-2 for various nationalities and use sensitivity analysis to explore how the number of internet users and the country's economic status impacts the sentiment of the stories. To reduce the propagation of biases through large language models (LLM), we explore the debiasing method of adversarial triggering. Our results show that GPT-2 demonstrates significant bias against countries with lower internet users, and adversarial triggering effectively reduces the same.

Classification margins are commonly used to estimate the generalization ability of machine learning models. We present an empirical study of these margins in artificial neural networks. A global estimate of margin size is usually used in the literature. In this work, we point out seldom considered nuances regarding classification margins. Notably, we demonstrate that some types of training samples are modelled with consistently small margins while affecting generalization in different ways. By showing a link with the minimum distance to a different-target sample and the remoteness of samples from one another, we provide a plausible explanation for this observation. We support our findings with an analysis of fully-connected networks trained on noise-corrupted MNIST data, as well as convolutional networks trained on noise-corrupted CIFAR10 data.

In supervised machine learning, the choice of loss function implicitly assumes a particular noise distribution over the data. For example, the frequently used mean squared error (MSE) loss assumes a Gaussian noise distribution. The choice of loss function during training and testing affects the performance of artificial neural networks (ANNs). It is known that MSE may yield substandard performance in the presence of outliers. The Cauchy loss function (CLF) assumes a Cauchy noise distribution, and is therefore potentially better suited for data with outliers. This papers aims to determine the extent of robustness and generalisability of the CLF as compared to MSE. CLF and MSE are assessed on a few handcrafted regression problems, and a real-world regression problem with artificially simulated outliers, in the context of ANN training. CLF yielded results that were either comparable to or better than the results yielded by MSE, with a few notable exceptions.

Recent work on predicting category structure with distributional models, using either static word embeddings (Heyman and Heyman, 2019) or contextualized language models (CLMs) (Misra et al., 2021), report low correlations with human ratings, thus calling into question their plausibility as models of human semantic memory. In this work, we revisit this question testing a wider array of methods for probing CLMs for predicting typicality scores. Our experiments, using BERT (Devlin et al., 2018), show the importance of using the right type of CLM probes, as our best BERT-based typicality prediction methods substantially improve over previous works. Second, our results highlight the importance of polysemy in this task: our best results are obtained when using a disambiguation mechanism. Finally, additional experiments reveal that Information Contentbased WordNet (Miller, 1995), also endowed with disambiguation, match the performance of the best BERT-based method, and in fact capture complementary information, which can be combined with BERT to achieve enhanced typicality predictions.

As a result, they are often brittle and unable to adapt to new domains, can treat individuals or subgroups unfairly, and have limited ability to explain their actions or recommendations [197, 235] reducing the trust of human users [118]. Following this, a new area of research, trustworthy AI, has recently received much attention from several policymakers and other regulatory organizations. The resulting guidelines (e.g., [184, 186, 187]), introduced to increase trust in AI systems, make developing trustworthy AI not only a technical (research) and social endeavor but also an organizational and (legal) obligational requirement. In this paper, we set out to demonstrate, through an extensive survey, that causal modeling and reasoning is an emerging and very useful tool for enabling current AI systems to become trustworthy. Causality is the science of reasoning about causes and effects. Cause-and-effect relationships are central to how we make sense of the world around us, how we act upon it, and how we respond to changes in our environment. In AI, research in causality was pioneered by the Turing award winner Judea Pearl long back in his 1995 seminal paper [194]. Since then, many researchers have contributed to the development of a solid mathematical basis for causality; see, for example, the books [79, 196, 201], the survey [90] and seminal papers [197, 235].

Ongoing progress in computational intelligence (CI) has led to an increased desire to apply CI techniques for the purpose of improving software engineering processes, particularly software testing. Existing state-of-the-art automated software testing techniques focus on utilising search algorithms to discover input values that achieve high execution path coverage. These algorithms are trained on the same code that they intend to test, requiring instrumentation and lengthy search times to test each software component. This paper outlines a novel genetic programming framework, where the evolved solutions are not input values, but micro-programs that can repeatedly generate input values to efficiently explore a software component's input parameter domain. We also argue that our approach can be generalised such as to be applied to many different software systems, and is thus not specific to merely the particular software component on which it was trained.

The existence of external (``side'') semantic knowledge has been shown to result in more expressive computational event models. To enable the use of side information that may be noisy or missing, we propose a semi-supervised information bottleneck-based discrete latent variable model. We reparameterize the model's discrete variables with auxiliary continuous latent variables and a light-weight hierarchical structure. Our model is learned to minimize the mutual information between the observed data and optional side knowledge that is not already captured by the new, auxiliary variables. We theoretically show that our approach generalizes past approaches, and perform an empirical case study of our approach on event modeling. We corroborate our theoretical results with strong empirical experiments, showing that the proposed method outperforms previous proposed approaches on multiple datasets.

Generating text with autoregressive language models (LMs) is of great importance to many natural language processing (NLP) applications. Previous solutions for this task often produce text that contains degenerative expressions or lacks semantic consistency. Recently, Su et al. introduced a new decoding method, contrastive search, based on the isotropic representation space of the language model and obtained new state of the art on various benchmarks. Additionally, Su et al. argued that the representations of autoregressive LMs (e.g. GPT-2) are intrinsically anisotropic which is also shared by previous studies. Therefore, to ensure the language model follows an isotropic distribution, Su et al. proposed a contrastive learning scheme, SimCTG, which calibrates the language model's representations through additional training. In this study, we first answer the question: "Are autoregressive LMs really anisotropic?". To this end, we extensively evaluate the isotropy of LMs across 16 major languages. Surprisingly, we find that the anisotropic problem only exists in the two specific English GPT-2-small/medium models. On the other hand, all other evaluated LMs are naturally isotropic which is in contrast to the conclusion drawn by previous studies. Based on our findings, we further assess the contrastive search decoding method using off-the-shelf LMs on four generation tasks across 16 languages. Our experimental results demonstrate that contrastive search significantly outperforms previous decoding methods without any additional training. More notably, on 12 out of the 16 evaluated languages, contrastive search performs comparably with human-level performances as judged by human evaluations. Our code and other related resources are publicly available at https://github.com/yxuansu/Contrastive_Search_Is_What_You_Need.

Activation functions can have a significant impact on reducing the topological complexity of input data and therefore improve the performance of the model. Selecting a suitable activation function is an essential step in neural model design. However, the choice of activation function is seldom discussed or explored in Transformer-based language models. Their activation functions are chosen beforehand and then remain fixed from pre-training to fine-tuning. As a result, the inductive biases they imposed on models cannot be adjusted during this long life cycle. Moreover, subsequently developed models (e.g., RoBERTa, BART, and GPT-3) often follow up prior work (e.g., BERT) to use the same activation function without justification. In this paper, we investigate the effectiveness of using Rational Activation Function (RAF), a learnable activation function, in the Transformer architecture. In contrast to conventional, predefined activation functions, RAFs can adaptively learn optimal activation functions during training according to input data. Our experiments show the RAF-based Transformer (RAFT) achieves a lower validation perplexity than a vanilla BERT with the GELU function. We further evaluate RAFT on downstream tasks in low- and full-data settings. Our results show that RAFT outperforms the counterpart model across the majority of tasks and settings. For instance, RAFT outperforms vanilla BERT on the GLUE benchmark by 5.71 points on average in low-data scenario (where 100 training examples are available) and by 2.05 points on SQuAD in full-data setting. Analysis of the shapes of learned RAFs further unveils that they substantially vary between different layers of the pre-trained model and mostly look very different from conventional activation functions. RAFT opens a new research direction for analyzing and interpreting pre-trained models according to the learned activation functions.