Natural Language Understanding seminars

Visual information is central to healthcare, agriculture, autonomous systems and communication, but remains expensive to encode, difficult to interpret in detail and vulnerable to manipulation. Manoranjan Paul presents research that combines domain knowledge, structural constraints and careful treatment of noisy observations to make visual AI dependable in practical settings. Four connected directions organize the talk. Compact representations of video and three-dimensional data can support several downstream tasks without separate encodings. Vision-language methods connect image components with semantic descriptions under limited supervision. Efficient task-specific video representations support reasoning across space and time and video question answering. Structural evidence in generated images helps distinguish synthetic content from authentic images. The presentation draws on digital health, environmental sensing and intelligent systems, and closes with research challenges in integrated multimodal pipelines and emerging approaches to computing for imaging.

Jie Yang presents BRIDGE, a multilingual evaluation benchmark built from real clinical tasks and more than one million samples derived from electronic health records. The benchmark addresses the gap between simplified examination-style tests and the complexity of clinical data, while enabling comparison of rapidly changing language models. Evaluation of 95 models involved over 24,000 experiments and 39 million predictions. Results vary substantially by model family, task and language, with some open-source systems matching proprietary models. The study also finds that chain-of-thought prompting frequently reduces accuracy on these tasks and examines stigmatizing language produced during model reasoning at large scale.

Large speech and language models offer testable algorithmic hypotheses about how the brain understands language. Laura Gwilliams examines neural recordings at several spatial scales, from population activity measured with MEG and intracranial EEG to individual neurons sampled across cortical depth with human Neuropixels probes. The studies ask which representations and operations connect speech sounds to meaning, and how information at different timescales is nested across time and cortical space. Comparing biological recordings with artificial models reveals candidate mechanisms for exchanging information across hierarchical structures and for implementing language understanding at the level of whole neural systems.

This webinar convened researchers at the intersection of Artificial Intelligence and Neuroscience to investigate how large language models (LLMs) can serve as valuable “model organisms” for understanding human language processing. Presenters showcased evidence that brain recordings (fMRI, MEG, ECoG) acquired while participants read or listened to unconstrained speech can be predicted by representations extracted from state-of-the-art text- and speech-based LLMs. In particular, text-based LLMs tend to align better with higher-level language regions, capturing more semantic aspects, while speech-based LLMs excel at explaining early auditory cortical responses. However, purely low-level features can drive part of these alignments, complicating interpretations. New methods, including perturbation analyses, highlight which linguistic variables matter for each cortical area and time scale. Further, “brain tuning” of LLMs—fine-tuning on measured neural signals—can improve semantic representations and downstream language tasks. Despite open questions about interpretability and exact neural mechanisms, these results demonstrate that LLMs provide a promising framework for probing the computations underlying human language comprehension and production at multiple spatiotemporal scales.

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Natural language understanding comprises a wide range of diverse tasks such as textual entailment, question answering, semantic similarity assessment, and document classification. Although large unlabeled text corpora are abundant, labeled data for learning these specific tasks is scarce, making it challenging for discriminatively trained models to perform adequately. We demonstrate that large gains on these tasks can be realized by generative pre-training of a language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task. In contrast to previous approaches, we make use of task-aware input transformations during fine-tuning to achieve effective transfer while requiring minimal changes to the model architecture. We demonstrate the effectiveness of our approach on a wide range of benchmarks for natural language understanding. Our general task-agnostic model outperforms discriminatively trained models that use architectures specifically crafted for each task, significantly improving upon the state of the art in 9 out of the 12 tasks studied. For instance, we achieve absolute improvements of 8.9% on commonsense reasoning (Stories Cloze Test), 5.7% on question answering (RACE), and 1.5% on textual entailment (MultiNLI).

Qualitative coding is the process of categorizing and labeling raw data to identify themes, patterns, and concepts within qualitative research. This process requires significant time, reflection, and discussion, often characterized by inherent subjectivity and uncertainty. Here, we explore the possibility to leverage large language models (LLM) to enhance the process and assist researchers with qualitative coding. LLMs, trained on extensive human-generated text, possess an architecture that renders them capable of understanding the broader context of a conversation or text. This allows them to extract patterns and meaning effectively, making them particularly useful for the accurate extraction and coding of relevant themes. In our current approach, we employed the chatGPT 3.5 Turbo API, integrating it into the qualitative coding process for data from the SWISS100 study, specifically focusing on data derived from centenarians' experiences during the Covid-19 pandemic, as well as a systematic centenarian literature review. We provide several instances illustrating how our approach can assist researchers with extracting and coding relevant themes. With data from human coders on hand, we highlight points of convergence and divergence between AI and human thematic coding in the context of these data. Moving forward, our goal is to enhance the prototype and integrate it within an LLM designed for local storage and operation (LLaMa). Our initial findings highlight the potential of AI-enhanced qualitative coding, yet they also pinpoint areas requiring attention. Based on these observations, we formulate tentative recommendations for the optimal integration of LLMs in qualitative coding research. Further evaluations using varied datasets and comparisons among different LLMs will shed more light on the question of whether and how to integrate these models into this domain.

Young children have sophisticated representations of their visual and linguistic environment. Where do these representations come from? How much knowledge arises through generic learning mechanisms applied to sensory data, and how much requires more substantive (possibly innate) inductive biases? We examine these questions by training neural networks solely on longitudinal data collected from a single child (Sullivan et al., 2020), consisting of egocentric video and audio streams. Our principal findings are as follows: 1) Based on visual only training, neural networks can acquire high-level visual features that are broadly useful across categorization and segmentation tasks. 2) Based on language only training, networks can acquire meaningful clusters of words and sentence-level syntactic sensitivity. 3) Based on paired visual and language training, networks can acquire word-referent mappings from tens of noisy examples and align their multi-modal conceptual systems. Taken together, our results show how sophisticated visual and linguistic representations can arise through data-driven learning applied to one child’s first-person experience.

Analogical reasoning –learning about new things by relating it to previous knowledge– lies at the heart of human intelligence and creativity and forms the core of educational practice. Children start creating and using analogies early on, making incredible progress moving from associative processes to successful analogical reasoning. For example, if we ask a four-year-old “Horse belongs to stable like chicken belongs to …?” they may use association and reply “egg”, whereas older children will likely give the intended relational response “chicken coop” (or other term to refer to a chicken’s home). Interestingly, despite state-of-the-art AI-language models having superhuman encyclopedic knowledge and superior memory and computational power, our pilot studies show that these large language models often make mistakes providing associative rather than relational responses to verbal analogies. For example, when we asked four- to eight-year-olds to solve the analogy “body is to feet as tree is to …?” they responded “roots” without hesitation, but large language models tend to provide more associative responses such as “leaves”. In this study we examine the similarities and differences between children's and six large language models' (Dutch/multilingual models: RobBERT, BERT-je, M-BERT, GPT-2, M-GPT, Word2Vec and Fasttext) responses to verbal analogies extracted from an online adaptive learning environment, where >14,000 7-12 year-olds from the Netherlands solved 20 or more items from a database of 900 Dutch language verbal analogies.