Amit Sharma (Mictosoft Research, India)

Talk title: Teaching Causal Reasoning to Transformers

Abstract: 

Large language models (LLMs) have demonstrated remarkable accuracy in identifying cause- and-effect relationships across various scientific fields. However, their ability to reason over these relationships remains inconsistent. In this work, we introduce an axiomatic approach to enhance causal reasoning in transformer models. By training on synthetic demonstrations of causal axioms, we show that even models with fewer than 100 million parameters can achieve reasoning capabilities comparable to much larger models, such as Phi and GPT-4. This suggests that axiomatic training offers a new method for imparting causal reasoning to LLMs, without the need for active interventions. Additionally, it may help explain how models like GPT-4, which are trained solely on observational data, can exhibit seemingly sophisticated reasoning.  

Rita Goldstein (Icahn School of Medicine at Mount Sinai, USA)

Talk title: Natural Language Processing of Spontaneous Speech in Drug Addiction: A Brain Biomarker of Treatment Effects and a Predictor of Outcomes 

Abstract: 

Escalating morbidity and mortality with opioid and stimulant use disorders makes it imperative to identify reliable and valid behavioral markers of brain function that change with treatment and predict outcomes. A naturalistic approach targeting spontaneous speech can serve as a readily accessible behavioral marker of neural plasticity with treatment, amenable for analyses with innovative artificial intelligence (AI), including natural language processing (NLP) approaches. Language, a ubiquitous, dense and measurable yet complex brain function that engages numerous higher-order executive functions (including the prefrontal cortex, PFC), is a distinctively human medium to express emotions, thoughts, and perceptions and to communicate. Concomitant with evidence for PFC recovery with treatment and abstinence in people with heroin and cocaine use disorders, using NLP our results show the impact of 3- months of treatment also on verbal movie-recall, further tracking the brain’s shared representation of this drug-themed complex stimulus, and its change with time; in contrast to common lab-based tasks, movies establish a dynamic narrative context, approximating real world processing by allowing a glimpse into the way groups of individuals with common addiction related experiences collectively attribute salience in their environment, addressing the social nature of drug use and recovery. Our results also show that NLP of speech about the consequences of drug use or abstinence outperformed numerous other models (including those incorporating drug use itself) as a predictor of 1-year outcomes, suggesting that hidden predictors of longitudinal craving, withdrawal, abstinence and drug use are embedded in peoples’ verbal outputs at baseline. The machine-learning community could only unlock the power of AI by first painstakingly curating large-scale training datasets. Using rich data that can be collected at scale, here we demonstrate a naturalistic ecologically valid approach to identify prognostic bio-markers of vulnerability and resilience in drug addiction. These bio-markers could ultimately facilitate the identification and design of neuroscience-informed empirically-based individually-tailored intervention and prevention approaches, timely deployed to improve outcomes, optimize recovery and save lives in people with drug addiction. 

Roger Levy (Department of Brain and Cognitive Sciences at MIT, USA)

Talk title: Cognitive Science of Language in the Era of Large Language Models 

Abstract: 

Large language models (LLMs) are the first human-created artifacts whose text processing and generation capabilities seem to approach our own. But the hardware they run on is vastly different than ours, and the software implementing them probably is too. How, then, can we use LLMs to advance the science of language in the human mind? In this talk I give examples of three key ways LLMs can contribute to the cognitive science of language: they can help us place lower bounds on the learnability of linguistic generalizations; they can help us reverse-engineer human language processing mechanisms; and they can help us develop hypotheses for the interface between language and other cognitive mechanisms. 

Hadas Raviv (Princeton Neuroscience Institute, USA)

Talk title: The First 1000 Days - Harnessing Real-World Data to Model the Complexity of Language Acquisition 

Abstract: 

Human motor, cognitive, and social skills undergo extraordinary development during life’s first 1000 days (1kD). Babies are born helpless, equipped with minimal cognitive-motor skills, and must rely on their caregivers to provide for their needs. However, within a few years, they walk, climb, talk, and reason about the world. Despite the large body of research devoted to studying the first 1000 days, our understanding of the developmental processes that shape children’s future outcomes is limited. One of the biggest impediments to modeling human development is the lack of dense measurements of children’s lives during their first few years of life. Until recently, capturing, managing, and analyzing large longitudinal datasets containing the rich sensory input with which the child is immersed has been technologically challenging.

In this talk, I will introduce the One Thousand Days (1kD) project. In this project we have been capturing the daily lives of 17 babies from 15 diverse families in the U.S. East Coast, recording their home environment 12 hours a day, 7 days a week, for a period of 1000 days. To date, we have amassed an extraordinary 230 years' worth of data, which is not only vast but also remarkably diverse and rich, encompassing many aspects of human behavior and interaction within real-life family environments.

The 1kD dataset offers an unprecedented opportunity to quantify, for the first time, the natural, everyday statistics within infants' environments that drive learning. It also enables the development and testing of human-centered computational models of language acquisition using ecologically valid, child-centered data. I will present our modeling framework for language acquisition during the first year of life, emphasizing the cognitive and biological factors that are crucial for creating human-centered models of language development as well as some preliminary results and findings.

Nitay Calderon (The Faculty of Data and Decision Sciences at the Technion, Israel)  

Talk title: The Role of NLP Interpretability in Advancing the Human Mind and Social Science Research  

Abstract: 

The role of language in modern science goes far beyond its use as a tool for communicating research, sharing knowledge, or exploring new ideas. Language is fundamental to understanding human behavior, thoughts, and social development, serving as a gateway to comprehending the human mind. Given this centrality, it is no surprise that recent advances in Natural Language Processing (NLP) and the rise of Large Language Models (LLMs) have led researchers to leverage NLP for modeling scientific phenomena, decoding complex patterns, and uncovering insights about humanity. In this talk, I will explore how NLP interpretability methods are instrumental for the NLP4Science approach. Specifically, interpretability enables us to understand the underlying mechanisms and patterns the NLP model identifies, facilitating deeper comprehension and advancing scientific discoveries. I will review key interpretability paradigms, compare their strengths and limitations, and showcase their application in diverse scientific disciplines. For example, probing techniques are commonly used in neuroscience by aligning NLP representations with brain signals, and causal NLP methods such as counterfactuals are used in political science. I will then present a comprehensive analysis of over two thousand interpretability papers, highlighting trends in NLP interpretability paradigms and comparing them to those in other fields. Finally, I will discuss the broader implications of these findings and offer recommendations for promoting the adoption of NLP interpretability techniques in fields beyond NLP.