The distribution of subpopulations is an important property hidden within a dataset. Uncovering and analyzing the subpopulation distribution within datasets provides a comprehensive understanding of the datasets, standing as a powerful tool beneficial to various downstream tasks, including Dataset Subpopulation Organization, Subpopulation Shift, and Slice Discovery. Despite its importance, there has been no work that systematically explores the subpopulation distribution of datasets to our knowledge. To address the limitation and solve all the mentioned tasks in a unified way, we introduce a novel concept of subpopulation structures to represent, analyze, and utilize subpopulation distributions within datasets. To characterize the structures in an interpretable manner, we propose the Subpopulation Structure Discovery with Large Language Models (SSD-LLM) framework, which employs world knowledge and instruction-following capabilities of Large Language Models (LLMs) to linguistically analyze informative image captions and summarize the structures. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery. Furthermore, we propose complete workflows to address downstream tasks, named Task-specific Tuning, showcasing the application of the discovered structure to a spectrum of subpopulation-related tasks, including dataset subpopulation organization, subpopulation shift, and slice discovery.

We reassess the Knowledge Neuron (KN) Thesis: an interpretation of the mechanism underlying the ability of large language models to recall facts from a training corpus. This nascent thesis proposes that facts are recalled from the training corpus through the MLP weights in a manner resembling key-value memory, implying in effect that "knowledge" is stored in the network. Furthermore, by modifying the MLP modules, one can control the language model's generation of factual information. The plausibility of the KN thesis has been demonstrated by the success of KN-inspired model editing methods (Dai et al., 2022; Meng et al., 2022). We find that this thesis is, at best, an oversimplification. Not only have we found that we can edit the expression of certain linguistic phenomena using the same model editing methods but, through a more comprehensive evaluation, we have found that the KN thesis does not adequately explain the process of factual expression. While it is possible to argue that the MLP weights store complex patterns that are interpretable both syntactically and semantically, these patterns do not constitute "knowledge." To gain a more comprehensive understanding of the knowledge representation process, we must look beyond the MLP weights and explore recent models' complex layer structures and attention mechanisms. Recent research has highlighted the remarkable ability of large pretrained language models (PLMs) to recall facts from a training corpus (Petroni et al., 2019). The underlying mechanism by which this information is stored and retrieved within PLMs, however, remains a subject of intensive investigation. The Knowledge Neuron (KN) Thesis has been recently proposed as a novel framework for interpreting language models (LMs) (Dai et al., 2022; Meng et al., 2022; 2023). This thesis suggests that LMs operate akin to key-value memories, recalling facts from the training corpus through the multi-layer perceptron (MLP) weights. Therefore, a significant implication of the KN thesis is that factual information generation by LMs can be controlled by modifying the MLP modules. Should this manipulation of factual information recall become feasible, it could lead to the development of language models that are more controllable, interpretable, and factually aligned. The plausibility of the KN thesis is demonstrated by the success of KN-inspired model-editing methods. Dai et al. (2022) argued that relational facts can be localised to a handful of 2-5 MLP neurons.

Herbert L. Roitblat ABSTRACT Artificial intelligence systems exhibit many useful capabilities, but they appear to lack understanding. This essay describes how we could go about constructing a machine capable of understanding. As John Locke (1689) pointed out words are signs for ideas, which we can paraphrase as thoughts and concepts. To understand a word is to know and be able to work with the underlying concepts for which it is an indicator. Understanding between a speaker and a listener occurs when the speaker casts his or her concepts into words and the listener recovers approximately those same concepts. Current models rely on the listener to construct any potential meaning. The diminution of behaviorism as a psychological paradigm and the rise of cognitivism provide examples of many experimental methods that can be used to determine whether and to what extent a machine might understand and to make suggestions about how that understanding might be instantiated. I know there are not words enough in any language to answer all the variety of ideas that enter into men's discourses and reasonings. But this hinders not but that when any one uses any term, he may have in his mind a determined idea, which he makes it the sign of, and to which he should keep it steadily annexed during that present discourse. John Locke 1689 Artificial intelligence systems exhibit many useful capabilities, but as has often been said, they lack "understanding," which would be a critical capability for general intelligence. The transformer architecture on which current systems are based takes one string of tokens and produces another string of tokens (one token at a time) based on the aggregated statistics of the associations among tokens. The representations mediating between the inputs (e.g., prompts) and their production is one purely of the statistical relations among the word tokens. In the case of large language models, we know these facts to be true because this is how the models were designed and they were trained on a kind of fill-in-the-blank test to guess the next word. What exactly would it mean for an artificial intelligence system to understand? How would we know that it does?

Large Language Models (LLMs) are deep learning models designed to generate text based on textual input. Although researchers have been developing these models for more complex tasks such as code generation and general reasoning, few efforts have explored how LLMs can be applied to combinatorial problems. In this research, we investigate the potential of LLMs to solve the Travelling Salesman Problem (TSP). Utilizing GPT-3.5 Turbo, we conducted experiments employing various approaches, including zero-shot in-context learning, few-shot in-context learning, and chain-of-thoughts (CoT). Consequently, we fine-tuned GPT-3.5 Turbo to solve a specific problem size and tested it using a set of various instance sizes. The fine-tuned models demonstrated promising performance on problems identical in size to the training instances and generalized well to larger problems. Furthermore, to improve the performance of the fine-tuned model without incurring additional training costs, we adopted a self-ensemble approach to improve the quality of the solutions.

The rapid proliferation of large language models and natural language processing (NLP) applications creates a crucial need for uncertainty quantification to mitigate risks such as hallucinations and to enhance decision-making reliability in critical applications. Conformal prediction is emerging as a theoretically sound and practically useful framework, combining flexibility with strong statistical guarantees. Its model-agnostic and distribution-free nature makes it particularly promising to address the current shortcomings of NLP systems that stem from the absence of uncertainty quantification. This paper provides a comprehensive survey of conformal prediction techniques, their guarantees, and existing applications in NLP, pointing to directions for future research and open challenges.

Sentiment analysis is one of the most widely used techniques in text analysis. Recent advancements with Large Language Models have made it more accurate and accessible than ever, allowing researchers to classify text with only a plain English prompt. However, "sentiment" entails a wide variety of concepts depending on the domain and tools used. It has been used to mean emotion, opinions, market movements, or simply a general ``good-bad'' dimension. This raises a question: What exactly are language models doing when prompted to label documents by sentiment? This paper first overviews how sentiment is defined across different contexts, highlighting that it is a confounded measurement construct in that it entails multiple variables, such as emotional valence and opinion, without disentangling them. I then test three language models across two data sets with prompts requesting sentiment, valence, and stance classification. I find that sentiment labels most strongly correlate with valence labels. I further find that classification improves when researchers more precisely specify their dimension of interest rather than using the less well-defined concept of sentiment. I conclude by encouraging researchers to move beyond "sentiment" when feasible and use a more precise measurement construct.

Generalization to unseen data is a key desideratum for deep networks, but its relation to classification accuracy is unclear. Using a minimalist vision dataset and a measure of generalizability, we show that popular networks, from deep convolutional networks (CNNs) to transformers, vary in their power to extrapolate to unseen classes both across layers and across architectures. Accuracy is not a good predictor of generalizability, and generalization varies non-monotonically with layer depth. In deep learning for classification, generalization is typically considered in the context of training versus test sets (Zhang et al., 2021), where both contain examples from the same set of classes, and is measured via test set accuracy. Our goal is to investigate the capacity of a network to generalize its classification power to similar classes absent in the training set (unseen).

Pathology reports are rich in clinical and pathological details but are often presented in free-text format. The unstructured nature of these reports presents a significant challenge limiting the accessibility of their content. In this work, we present a practical approach based on the use of large multimodal models (LMMs) for automatically extracting information from scanned images of pathology reports with the goal of generating a standardised report specifying the value of different fields along with estimated confidence about the accuracy of the extracted fields. The proposed approach overcomes limitations of existing methods which do not assign confidence scores to extracted fields limiting their practical use. The proposed framework uses two stages of prompting a Large Multimodal Model (LMM) for information extraction and validation. The framework generalises to textual reports from multiple medical centres as well as scanned images of legacy pathology reports. We show that the estimated confidence is an effective indicator of the accuracy of the extracted information that can be used to select only accurately extracted fields. We also show the prognostic significance of structured and unstructured data from pathology reports and show that the automatically extracted field values significant prognostic value for patient stratification. The framework is available for evaluation via the URL: https://labieb.dcs.warwick.ac.uk/.

'The Big Weekend Show' analyzes the possibilities of artificial intelligence when it comes to influencing voters. ChatGPT has changed my life -- and yours, even if you don't use it as much as I do. You've probably noticed the new AI search bar in all the Meta apps, including Facebook and Instagram. It won't be long before all your most-used apps and services integrate chatbots. I'm giving one away to someone who tries my free daily tech newsletter.

Emails from the Department of Justice's antitrust case against Google revealed how Microsoft executives were alarmed by and even envious of Google's AI lead. In an email thread, CTO Kevin Scott wrote he was "very, very worried" about Google's rapidly growing AI capabilities. He said he initially dismissed the company's "game-playing stunts," likely referring to Google's AlphaGo models. The emails reference Gmail's autocomplete features, which execs called "scary good." Microsoft struggled to copy Google's BERT-large, an AI model that deciphers the meaning and context of words in a sentence.