Recent rapid progress in the creation of artificial intelligence (AI) systems has been driven in large part by innovations in architectures and algorithms for developing large scale artificial neural networks. As a consequence, it’s natural to ask what role abstract principles of intelligence — such as Bayes’ rule — might play in developing intelligent machines. In this talk, I will argue that there is a new way in which Bayes can be used in the context of AI, more akin to how it is used in cognitive science: providing an abstract description of how agents should solve certain problems and hence a tool for understanding their behavior. This new role is motivated in large part by the fact that we have succeeded in creating intelligent systems that we do not fully understand, making the problem for the machine learning researcher more closely parallel that of the cognitive scientist. I will talk about how this perspective can help us think about making machines with better informed priors about the world and give us insight into their behavior by directly creating cognitive models of neural networks.

Bayes in the age of intelligent machines

For nearly a decade now, the problem that has been top of mind for me is how we might enable AI systems to robustly optimize for what people want, and to avoid causing harm – from robots and self-driving cars, to assistive devices and deep brain stimulation, to theory and toy models, to large language models and now Gemini. In this talk, I’ll take the opportunity to share a bit about my journey in this space, what lessons I’ve learned, and how we’re approaching the safety and alignment of frontier models at Google DeepMind.

My Journey in AI Safety and Alignment

As capabilities of foundation models skyrocket, openness plummets. In this talk, I argue that open-source models are essential for the long-term goal of building a rigorous foundation for AI. Greater access—from API to open-weight to open-source—enables deeper forms of research. API access allows us to push the frontier of agents, and I will present our recent work on simulation and problem-solving agents. Open weights enables reproducible research on safety, interpretability, and more generally, “model forensics”. Open-source unlocks fundamental innovations in architectures, training procedures, and data curation methods. Of course, the key obstacle for building open-source models is the resources required (data, compute, and research/engineering). I will conclude with some promising directions that leverage the community that bring us closer to the vision of open-source foundation models.

Open-Source and Science in the Era of Foundation Models

NLP research made significant progress, and our community’s achievements are becoming deeply integrated in society. The recent paradigm shift due to rapid advances in Large Language Models (LLMs) offers immense potential, but also led NLP to become more homogeneous. In this talk, I will argue for the importance of embracing variation in research, which will lead to more innovation, and in turn, trust. I will give an overview of current challenges and show how they led to the loss of trust in our models. To counter this, I propose to embrace variation in three key areas: inputs to models, outputs of models and research itself. Embracing variation holistically will be crucial to move our field towards more trustworthy human-facing NLP.

Are LLMs Narrowing Our Horizon? Let’s Embrace Variation in NLP!

Large Language Models (LLMs) are on track to reverse what seemed like an inexorable shift of AI from explicit to tacit knowledge tasks. Trained as they are on everything ever written on the web, LLMs exhibit “approximate omniscience”–they can provide answers to all sorts of queries, but with nary a guarantee. This could herald a new era for knowledge-based AI systems–with LLMs taking the role of (blowhard?) experts. But first, we have to stop confusing the impressive style/form of the generated knowledge for correct/factual content, and resist the temptation to ascribe reasoning, planning, self-critiquing etc. powers to approximate retrieval by these n-gram models on steroids. We have to focus instead on LLM-Modulo techniques that complement the unfettered idea generation of LLMs with careful vetting by model-based verifiers (the models underlying which themselves can be teased out from LLMs in semi-automated fashion). In this talk, I will reify this vision and attendant caveats in the context of our ongoing work on understanding the role of LLMs in planning tasks.

Can LLMs Reason and Plan?

 Adapting a model trained on vast amounts of data to new tasks with limited labeled data has long been a challenging problem, and over the years, a diverse range of techniques have been explored. Effective model adaptation requires achieving high accuracy through task-specific specialization without forgetting previous learnings, robustly handling the high variance from limited task-relevant supervision, and doing so efficiently with minimal compute and memory overheads. Recently, large language models (LLMs) have demonstrated remarkable ease of adaptation to new tasks with just a few examples provided in context, without any explicit training for such a capability. Puzzled by this apparent success, many researchers have sought to explain why in-context learning (ICL) works, but we still have only an incomplete understanding. In this talk, we examine this emerging phenomenon and assess its potential to meet our longstanding model adaptation goals in terms of accuracy, robustness, and efficiency.

Does In-Context-Learning Offer the Best Tradeoff in Accuracy, Robustness, and Efficiency for Model Adaptation?

Large language models motivate an approach to meaning known as distributional semantics, that words mean what they do because of how they’re distributed in language. In this talk I will describe some evidence from my lab that suggests metrics from large language models do a good job of predicting behavioral and neural responses to some aspects of human language. I go on to describe some research that highlights important differences in meaning processing in humans and the ‘understanding’ displayed by language models. Discrepancies are particularly noteworthy in studies of joke comprehension.

Distributional Semantics: What do large language models have to say?

How can Large Language Models (LLMs) and modern NLP be used to increase the use and the documentation of Indigenous languages which are in danger of disappearing? First, I report on the development of high-quality translators for Indigenous languages by fine-tuning SOTA machine translators with tiny amounts of data, and discuss how to avoid some common pitfalls. Next, I present prototypes built with Indigenous communities aiming to stimulate and facilitate writing, using LLM models to create spell-checkers, next-word predictors, and similar tools. Finally, I discuss a future for documentation where dying languages are preserved as interactive language models.

Harnessing the Power of LLMs to Vitalize Indigenous Languages

Recent years have witnessed the rise of increasingly larger and more sophisticated language models (LMs) capable of performing every task imaginable, sometimes at (super)human level. In this talk, I will argue that there is still space for specialist models in today’s NLP landscape. Such models can be dramatically more efficient, inclusive, and explainable. I will focus on two examples, opinion summarization and crosslingual semantic parsing and show how these two seemingly unrelated tasks can be addressed by explicitly learning task-specific representations. I will show how such representations can be further structured to allow search and retrieval, evidence-based generation, and cross-lingual alignment. Finally, I will discuss why we need to to use LLMs for what they are good at and remove the need for them to do things that can be done much better by smaller models.

Prompting is not all you need! Or why Structure and Representations still matter in NLP

That the recent LLM breakthroughs are solely due to scaling is a myth. Many difficult research problems had to be solved to make models like GPT4 and Mixtral possible. One of those difficult research problems is data quality. Data quality is a great challenge for NLP researchers with many opportunities for innovation and impact on current generative AI developments. I will focus on two examples in my talk: quality data for training a highly multilingual language model and quality data for instruction tuning.

Quality Data for LLMs: Challenges and Opportunities for NLP

**Karen Spärck Jones Award Lecture**<br>
The advent of generative AI technology produces transformative impact on the content creation landscape, offering alternative approaches to produce diverse, good-quality content across media, thereby reshaping the ecosystems of online content creation and publishing, but also raising concerns about market over-saturation and the potential marginalization of human creativity. Our recent work introduces a competition model generalized from the Tullock contest to analyze the tension between human creators and generative AI. Our theory and simulations suggest that despite challenges, a stable equilibrium between human and AI-generated content is possible. Our work contributes to understanding the competitive dynamics in the content creation industry, offering insights into the future interplay between human creativity and technological advancements in generative AI.

Underline digital video library