2024 Symposium

Yejin Choi is Wissner-Slivka Professor and a MacArthur Fellow at the Paul G. Allen School of Computer Science & Engineering at the University of Washington. She is also a senior director at AI2 overseeing the project Mosaic and a Distinguished Research Fellow at the Institute for Ethics in AI at the University of Oxford. Her research investigates if (and how) AI systems can learn commonsense knowledge and reasoning, if machines can (and should) learn moral reasoning, and various other problems in NLP, AI, and Vision including neuro-symbolic integration, language grounding with vision and interactions, and AI for social good. She is a co-recipient of 2 Test of Time Awards (at ACL 2021 and ICCV 2021), 8 Best/Outstanding Paper Awards (at ACL 2023, EMNLP 2023, NAACL 2022, ICML 2022, NeurIPS 2021, AAAI 2019, and ICCV 2013), the Borg Early Career Award (BECA) in 2018, the inaugural Alexa Prize Challenge in 2017, and IEEE AI's 10 to Watch in 2016.

Possible Impossibilities, Impossible Possibilities, and Paradoxes

Abstract: Generative AI has led to an unprecedented amount of global attention---both excitements and concerns, in part due to our relatively limited understanding about intelligence---both artificial and natural. In this talk, I will question if there can be possible impossibilities of large language models (i.e., the fundamental limits of transformers, if any) and the impossible possibilities of language models (i.e., seemingly impossible alternative paths beyond scale, if at all). I will then discuss a few paradoxes including the Generative AI Paradox: for AI, at least in its current form, generation capability may often exceed understanding capability, in stark contrast to human intelligence where generation (of e.g., novels, paintings) can be substantially harder than understanding.

Melanie Mitchell is a Professor at the Santa Fe Institute. Her current research focuses on conceptual abstraction, analogy-making, and visual recognition in artificial intelligence systems. 

Melanie is the author or editor of six books and numerous scholarly papers in the fields of artificial intelligence, cognitive science, and complex systems. Her book Complexity: A Guided Tour (Oxford University Press) won the 2010 Phi Beta Kappa Science Book Award and was named by Amazon.com as one of the ten best science books of 2009. Her latest book is Artificial Intelligence: A Guide for Thinking Humans (Farrar, Straus, and Giroux). 

Melanie originated the Santa Fe Institute's Complexity Explorer platform, which offers online courses and other educational resources related to the field of complex systems. Her online course “Introduction to Complexity” has been taken by over 25,000 students, and is one of Course Central’s “top fifty online courses of all time”.

Do Large Language Models Understand the World?

Abstract: I will survey a current, heated debate in the artificial intelligence (AI) research community on whether large pre-trained language models can be said to "understand" language — and the physical and social situations language encodes — in any important sense. I will describe arguments that have been made for and against such understanding, and, more generally, will discuss what methods can be used to fairly evaluate understanding and intelligence in AI systems. I will conclude with key questions for the broader sciences of intelligence that have arisen in light of these discussions.

Raphaël Millièrea is a Lecturer (Assistant Professor) in philosophy at Macquarie University in Sydney, Australia.

Raphaël's interests lie mainly within the philosophy of artificial intelligence, cognitive science, and mind. Particularly in gaining a better understanding of modern artificial neural networks based on deep learning architectures, such as large language models. Assessing the capacities and limitations of these models is a challenging and interdisciplinary endeavor. Philosophy can provide important insights to establish fair and meaningful comparisons between machines and humans in various domains, including language processing and reasoning. In return, studying artificial neural networks might inform theorizing about human cognition.

Raphaël is also interested in the nature and scope of self-representation; ethical and safety issues related to the development and deployment of AI; issues related to the relationship between art and technology; issues related to the conceptualization and taxonomy of global states of consciousness, such as post-comatose disorders of consciousness and drug-induced states; and methodological issues regarding the use of first-person reports in neuroscience.

Can we understand large language models?

Abstract: Large language models are perplexing. They exhibit human-like proficiency on sophisticated tasks, yet also fail spectacularly on others. Those who are impressed by their success believe they are stepping stones towards general artificial intelligence; those who are impressed by their failures see them as little more than convincing mimics. This has sparked a heated debate about the kind of cognitive capacities we can meaningfully attribute to these systems – if any. I will argue that focusing exclusively on the behavior of language models is a dead end to settle this dispute. Merely observing what language models can and cannot do only takes us so far; we need to understand how they do it. More specifically, we need to understand the causal mechanisms that drive their behavior. In principle, we can make progress towards this explanatory goal through careful interventions designed to reveal the internal causal structure of language models. In practice, however, this is a challenging project fraught with a number of potential theoretical and methodological pitfalls. After reviewing some of these challenges, I will conclude by discussing how gaining a mechanistic understanding of language model could help us settle deep philosophical questions about their cognitive capacities, and about the nature of cognition itself.

Paul Smolensky is a partner researcher in the Micosoft Deep Learning Group and part-year Krieger-Eisenhower Professor of Cognitive Science at Johns Hopkins University.His work focuses on the integration of symbolic and neural network computation for modeling reasoning and, especially, grammar in the human mind/brain. This work created: Harmony Networks (a.k.a. Restricted Boltzmann Machines); Tensor Product Representations; Optimality Theory and Harmonic Grammar (grammar frameworks grounded in neural computation); and Gradient Symbolic Computation. The work up through the early 2000’s is presented in the 2-volume MIT Press book with G Legendre, The Harmonic Mind.

Before assuming his position at Johns Hopkins he was a professor in the Computer Science Department of the University of Colorado at Boulder. Prior to that, as a postdoc and Research Professor, with G Hinton, D Rumelhart & J McClelland he was one of the founding members of the Parallel Distributed Processing Research Group at UCSD, which produced the bible of the previous wave of neural network modeling, the two-volume ‘PDP books’. His BA and PhD are in (Mathematical) Physics from Harvard and Indiana University.

He received the 2005 David E. Rumelhart Prize for Outstanding Contributions to the Formal Analysis of Human Cognition.

The current compositionality crisis in cognitive science

Abstract: Compositionality is key to our theories of language (and much of cognition). This means that encodings of information in the mind — mental representations — are composed of abstract parts which are arranged to form structures defined by the abstract relations holding among their parts; and given such a structure as input, the result of mental processing is another structure composed of the results of processing the input’s parts. The importance of compositionality is not annulled by the facts that it holds only approximately, that parts and structures are not purely discrete, and that it is far from exceptionless.Long-influential classic critiques argued that neural networks are severely limited by being inherently unsuited for compositionality. Current artificial intelligence results have turned this situation entirely upside down, creating a central crisis in theories of cognition. Neural networks of language, for example, strongly out-perform classic symbolic language-processing systems which are inherently compositional. The classic critiques have new importance because they now utterly fail to explain the unparalleled success of neural networks in AI.

In this talk we examine the classic premise that neural networks are inherently unsuited for compositionality and see how modern neural-network language-generating models such as Chat-GPT are in fact well-designed for a certain type of compositional processing.