Search for: All records

In this paper, Kirchenbauer et. al. use a novel watermarking technology to watermark the output of large language models (LLMs) like ChatGP, which is often in the form of AI-generated text, and mitigate the harms associated with the increasing usage of these technologies. They note some of the capabilities of these LLM models as writing documents, creating executable code, and answering questions, often with human-like capabilities. In addition, they list some of the harms as social engineering and election manipulation campaigns that exploit automated bots on social media platforms, creation of fake news and web content, and use of AI systems for cheating onacademic writing and coding assignments. As for implications for policy makers, this technology can be utilized as a means to regulate and oversee the use of these LLMs on all public and social fronts where their AI-generated text output could pose a potential harm, such as those listed by the authors. (Methods and Metrics, watermarking LLM output) 

Machine Translation (MT) remains one of the last NLP tasks where large language models (LLMs) have not yet replaced dedicated supervised systems. This work exploits the complementary strengths of LLMs and supervised MT by guiding LLMs to automatically post-edit MT with external feedback on its quality, derived from Multidimensional Quality Metric (MQM) annotations. Working with LLaMA-2 models, we consider prompting strategies varying the nature of feedback provided and then fine-tune the LLM to improve its ability to exploit the provided guidance. Through experiments on Chinese-English, English-German, and English-Russian MQM data, we demonstrate that prompting LLMs to post-edit MT improves TER, BLEU and COMET scores, although the benefits of fine-grained feedback are not clear. Fine-tuning helps integrate fine-grained feedback more effectively and further improves translation quality based on both automatic and human evaluation. 

Emphasizing problem formulation in AI literacy activities with children is vital, yet we lack empirical studies on their structure and affordances. We propose that participatory design involving teachable machines facilitates problem formulation activities. To test this, we integrated problem reduction heuristics into storyboarding and invited a university-based intergenerational design team of 10 children (ages 8-13) and 9 adults to co-design a teachable machine. We find that children draw from personal experiences when formulating AI problems; they assume voice and video capabilities, explore diverse machine learning approaches, and plan for error handling. Their ideas promote human involvement in AI, though some are drawn to more autonomous systems. Their designs prioritize values like capability, logic, helpfulness, responsibility, and obedience, and a preference for a comfortable life, family security, inner harmony, and excitement as end-states. We conclude by discussing how these results can inform the design of future participatory AI activities. 

In intelligent IoT networks, an IoT user is capable of sensing the spectrum and learning from its observation to dynamically access the wireless channels without interfering with the primary user’s signal. The network, however, is potentially subject to primary user emulation and jamming attacks. In the existing works, various attacks and defense mechanisms for spectrum sharing in IoT networks have been proposed. This paper systematically conducts a targeted survey of these efforts and proposes new approaches for future studies to strengthen the communication of IoT users. Our proposed methods involve the development of intelligent IoT devices that go beyond existing solutions, enabling them not only to share the spectrum with licensed users but also to effectively thwart potential attackers. First, considering practical aspects of imperfect spectrum sensing and delay, we propose to utilize online machine learning-based approaches to design spectrum sharing attack policies. We also investigate the attacker’s channel observation/sensing capabilities to design attack policies using time-varying feedback graph models. Second, taking into account the IoT devices’ practical characteristics of channel switching delay, we propose online learning-based channel access policies for optimal defense by the IoT device to guarantee the maximum network capacity. We then highlight future research directions, focusing on the defense of IoT devices against adaptive attackers. Finally, aided by concepts from intelligence and statistical factor analysis tools, we provide a workflow which can be utilized for devices’ intelligence factors impact analysis on the defense performance. 

AI systems have been known to amplify biases in real-world data. Explanations may help human-AI teams address these biases for fairer decision-making. Typically, explanations focus on salient input features. If a model is biased against some protected group, explanations may include features that demonstrate this bias, but when biases are realized through proxy features, the relationship between this proxy feature and the protected one may be less clear to a human. In this work, we study the effect of the presence of protected and proxy features on participants’ perception of model fairness and their ability to improve demographic parity over an AI alone. Further, we examine how different treatments—explanations, model bias disclosure and proxy correlation disclosure—affect fairness perception and parity. We find that explanations help people detect direct but not indirect biases. Additionally, regardless of bias type, explanations tend to increase agreement with model biases. Disclosures can help mitigate this effect for indirect biases, improving both unfairness recognition and decision-making fairness. We hope that our findings can help guide further research into advancing explanations in support of fair human-AI decision-making. 

This paper reviews the developmental literature on decision making, discussing how increased reliance on gist thinking explains the surprising finding that important cognitive biases increase from childhood to adulthood. This developmental trend can be induced experimentally by encouraging verbatim (younger) versus gist (older) ways of thinking. We then build on this developmental literature to assess the developmental stage of artificial intelligence (AI) and how its decision making compares with humans, finding that popular models are not only irrational but they sometimes resemble immature adolescents. To protect public safety and avoid risk, we propose that AI models build on policy frameworks already established to regulate other immature decision makers such as adolescents.

 

Academic tabular benchmarks often contain small sets of curated features. In contrast, data scientists typically collect as many features as possible into their datasets, and even engineer new features from existing ones. To prevent over-fitting in subsequent downstream modeling, practitioners commonly use automated feature selection methods that identify a reduced subset of informative features. Existing benchmarks for tabular feature selection consider classical downstream models, toy synthetic datasets, or do not evaluate feature selectors on the basis of downstream performance. We construct a challenging feature selection benchmark evaluated on downstream neural networks including transformers, using real datasets and multiple methods for generating extraneous features. We also propose an input-gradient-based analogue of LASSO for neural networks that outperforms classical feature selection methods on challenging problems such as selecting from corrupted or second-order features. 

Data attribution methods play a crucial role in understanding machine learning models, providing insight into which training data points are most responsible for model outputs during deployment. However, current state-of-the-art approaches require a large ensemble of as many as 300,000 models to accurately attribute model predictions. These approaches therefore come at a high computational cost, are memory intensive, and are hard to scale to large models or datasets. In this work, we focus on a minimalist baseline, utilizing the feature space of a backbone pretrained via self-supervised learning to perform data attribution. Our method is model-agnostic and scales easily to large datasets. We show results on CIFAR-10 and ImageNet, achieving strong performance that rivals or outperforms state-of-the-art approaches at a fraction of the compute or memory cost. Contrary to prior work, our results reinforce the intuition that a model's prediction on one image is most impacted by visually similar training samples. Our approach serves as a simple and efficient baseline for data attribution on images. 

This paper proposes solutions to detecting and mitigating the blatant replication and memorization of data used to train text-to-image generators, especially Stable Diffusion. The potential for diffusion models to reproduce copyrighted or private images without user knowledge poses significant ethical and legal challenges. For lawmakers, this highlights the need for clear guidelines and regulations around the use of such models, especially in commercial applications. 

The strength of modern generative models lies in their ability to be controlled through text-based prompts. Typical "hard" prompts are made from interpretable words and tokens, and must be hand-crafted by humans. There are also "soft" prompts, which consist of continuous feature vectors. These can be discovered using powerful optimization methods, but they cannot be easily interpreted, re-used across models, or plugged into a text-based interface. We describe an approach to robustly optimize hard text prompts through efficient gradient-based optimization. Our approach automatically generates hard text-based prompts for both text-to-image and text-to-text applications. In the text-to-image setting, the method creates hard prompts for diffusion models, allowing API users to easily generate, discover, and mix and match image concepts without prior knowledge on how to prompt the model. In the text-to-text setting, we show that hard prompts can be automatically discovered that are effective in tuning LMs for classification.