Counting the Bugs in ChatGPT's Wugs: A Multilingual Investigation into the Morphological Capabilities of a Large Language Model

Large language models (LLMs) have recently reached an impressive level of linguistic capability, prompting comparisons with human language skills. However, there have been relatively few systematic inquiries into the linguistic capabilities of the latest generation of LLMs, and those studies that do exist (i) ignore the remarkable ability of humans to generalize, (ii) focus only on English, and (iii) investigate syntax or semantics and overlook other capabilities that lie at the heart of human language, like morphology. Here, we close these gaps by conducting the first rigorous analysis of the morphological capabilities of ChatGPT in four typologically varied languages (specifically, English, German, Tamil, and Turkish). We apply a version of Berko's (1958) wug test to ChatGPT, using novel, uncontaminated datasets for the four examined languages. We find that ChatGPT massively underperforms purpose-built systems, particularly in English. Overall, our results -- through the lens of morphology -- cast a new light on the linguistic capabilities of ChatGPT, suggesting that claims of human-like language skills are premature and misleading.

Manifold-based Test Generation for Image Classifiers

Neural networks used for image classification tasks in critical applications must be tested with sufficient realistic data to assure their correctness. To effectively test an image classification neural network, one must obtain realistic test data adequate enough to inspire confidence that differences between the implicit requirements and the learned model would be exposed. This raises two challenges: first, an adequate subset of the data points must be carefully chosen to inspire confidence, and second, the implicit requirements must be meaningfully extrapolated to data points beyond those in the explicit training set. This paper proposes a novel framework to address these challenges. Our approach is based on the premise that patterns in a large input data space can be effectively captured in a smaller manifold space, from which similar yet novel test cases---both the input and the label---can be sampled and generated. A variant of Conditional Variational Autoencoder (CVAE) is used for capturing this manifold with a generative function, and a search technique is applied on this manifold space to efficiently find fault-revealing inputs. Experiments show that this approach enables generation of thousands of realistic yet fault-revealing test cases efficiently even for well-trained models.

Input Prioritization for Testing Neural Networks

Deep neural networks (DNNs) are increasingly being adopted for sensing and control functions in a variety of safety and mission-critical systems such as self-driving cars, autonomous air vehicles, medical diagnostics, and industrial robotics. Failures of such systems can lead to loss of life or property, which necessitates stringent verification and validation for providing high assurance. Though formal verification approaches are being investigated, testing remains the primary technique for assessing the dependability of such systems. Due to the nature of the tasks handled by DNNs, the cost of obtaining test oracle data---the expected output, a.k.a. label, for a given input---is high, which significantly impacts the amount and quality of testing that can be performed. Thus, prioritizing input data for testing DNNs in meaningful ways to reduce the cost of labeling can go a long way in increasing testing efficacy. This paper proposes using gauges of the DNN's sentiment derived from the computation performed by the model, as a means to identify inputs that are likely to reveal weaknesses. We empirically assessed the efficacy of three such sentiment measures for prioritization---confidence, uncertainty, and surprise---and compare their effectiveness in terms of their fault-revealing capability and retraining effectiveness. The results indicate that sentiment measures can effectively flag inputs that expose unacceptable DNN behavior. For MNIST models, the average percentage of inputs correctly flagged ranged from 88% to 94.8%.