2020

Keith Holyoak, Department of Psychology, UCLA
Friday, October 23, 2020

Abstract Semantic Relations in Mind, Brain, and Machines

Abstract semantic relations (e.g., category membership, part-whole, antonymy, cause-effect) are central to human intelligence, underlying the distinctively human ability to reason by analogy. I will describe a computational project (Bayesian Analogy with Relational Transformations) that aims to extract explicit representations of abstract semantic relations from non-relational inputs automatically generated by machine learning. BART’s representations predict patterns of typicality and similarity for semantic relations, as well as similarity of neural signals triggered by semantic relations during analogical reasoning. In this approach, analogy emerges from the ability to learn and compare relations; mapping emerges later from the ability to compare patterns of relations.

Recorded Talk

Richard Futrell, Department of Language Science, UC Irvine
Friday, September 18, 2020

Efficiency-based models of natural language: Predicting word order universals Using information theory

Why is human language the way it is? I claim that human languages can be modeled as systems for efficient communication: that is, codes that maximize information transfer subject to constraints on the cognitive resources used during language production and comprehension. I use this efficiency-based framework to formulate quantitative theories of word order, aiming to explain the cross-linguistic universals of word order documented by linguists as well as the statistical distribution of word orders in massively cross-linguistic corpus studies.

I present three results. First, I show that word orders in 54 languages are shaped by dependency locality: a pressure for words linked in syntactic dependencies to be close to each other, which minimizes working memory usage during language processing. Second, I introduce a new model of information-processing difficulty in online language processing, which simultaneously captures effects of probabilistic expectations and working memory constraints, recovering dependency locality as a special case and making new predictions, in particular about adjective order. Third, I present a computational framework in which grammars can be directly optimized for efficiency. When grammars are optimized to maximize information transfer while minimizing processing difficulty, they end up reproducing 8 typological universals of word order.

Recorded Talk

2019

Stuart Russell, Computer science, UC Berkeley
Friday, October 18, 2019

What if we succeed?

It is reasonable to expect that AI capabilities will eventually exceed those of humans across a range of real-world-decision making scenarios. Should this be a cause for concern, as Elon Musk, Stephen Hawking, and others have suggested? While some in the mainstream AI community dismiss the issue, I will argue instead that a fundamental reorientation of the field is required. The “standard model” in AI and related disciplines aims to create systems that optimize arbitrary objectives. As we shift to a regime in which systems are more capable than human beings, this model fails catastrophically. Instead, we need to learn how to build systems that will, in fact, be beneficial for us. I will show that it is useful to imbue systems with explicit uncertainty concerning the true objectives of the humans they are designed to help. This uncertainty causes machine and human behavior to be inextricably (and game-theoretically) linked, while opening up many new avenues for research.

David Freedman, Neurobiology, The University of Chicago
Friday, November 8, 2019

Neural circuits of cognition in artificial and biological neural networks

Humans and other advanced animals have a remarkable ability to interpret incoming sensory stimuli and plan task-appropriate behavioral responses. This talk will present parallel experimental and computational approaches aimed at understanding how visual feature encoding in upstream sensory cortical areas is transformed across the cortical hierarchy into more flexible task-related encoding in the parietal and prefrontal cortices. The experimental studies utilize multielectrode recording approaches to monitor activity of neuronal populations, as well as reversible cortical inactivation approaches, during performance of visual decision making tasks. In parallel, our computational work employs machine learning approaches to train recurrent artificial neural networks to perform the same tasks as in the experimental studies, allowing a deep investigation of putative neural circuit mechanisms used by both artificial and biological networks to solve cognitively demanding behavioral tasks.