Patrick Shafto, PhD - US grants

Research has shown that both children and adults differentiate between pedagogical and non-pedagogical situations, that pedagogical inferences override factual information and impede discovery learning in children, and that pedagogical demonstrations lead to predictable patterns of inferences in undergraduates. The purpose of this CAREER project is to explore the extent to which social-pedagogical context affects learning with preschool students and undergraduates. The project will contrast learning from a knowledgeable teacher with learning from a naive teacher, to explore how different inferences result from these situations, even when the exact same examples are provided.

The principal investigator proposes to use a Bayesian model of social-pedagogical learning across six laboratory-based studies to test the main hypothesis that pedagogical situations affect the inferences that learners draw from data in ways that are manipulable and predictable. For each of the six experimental studies the researcher will recruit approximately 50 preschoolers or undergraduates and randomly assign them to one of two conditions in a laboratory setting. The research will explore how inferences are drawn from a set of examples while manipulating how those examples are generated (e.g., from a knowledgeable teacher vs. at random). The goal of the project is to learn more about when, why, and how social-pedagogical context affects learning to inform a theoretical framework for understanding children and adult abilities to transmit and accumulate knowledge.

This INSPIRE award is partially funded by the Information Integration and Informatics program in the Division of Information and Intelligent Systems in the Directorate for Computer & Information Science & Engineering, the Perception, Action & Cognition program in the Division of Behavioral and Cognitive Sciences in the Directorate for Social, Behavioral & Economic Sciences, and the Office of Integrative Activities in the Office of the Director.

One of the most common uses of machine learning is to learn to replicate human decisions, a common example is recommender systems. In these systems, computers are trained to replicate the recommendation a collaboration of hundreds or thousands of humans would give, if that were possible. Most of the data used to train these systems are not from a controlled random sample, but are obtained from users based on outputs of algorithms (e.g., which search engine results do users click on?), which introduces bias into the process and ultimately impacts the quality of the results. This project addresses this problem by examining how the human decision process that creates these data in the first place is affected by the data coming from machine algorithms, how this in turn impacts the algorithms themselves, and how to ultimately adjust for human bias in the machine learning process. Specific areas tackled are filtering (e.g., web search) and recommender systems. The deep research into how the human decision process affects machine learning, and how machine learning impacts the human decision process, can provide significant advances in the accuracy and utility of systems using machine learning.

The project builds on analysis of machine learning algorithms based on Hidden Markov Models (HMMs). The formal analysis initially looks at "blind spots" - the impact of bias from users not getting complete (or a random sample) of data. Further analysis will be based on the outcome of two human experiments: Two category recommendation (labeling items, with items to be labelled chosen by random, active learning, and filter-based algorithms), and movie recommendation. The results will be used to develop improved machine learning approaches based on antidotes (altering learned models to reduce bias) and reactive learning (active learning that takes into account the human and machine biases). The PIs also have plans to capitalize on the lessons learned by providing examples of the use of cognitive science in a Web Mining course, and of the impact of machine learning in Data Science for Psychologists courses.

This Science of Learning Collaborative Network brings together psychology, education, and computer science researchers from Rutgers University, Temple University, Boston University, University of California-Berkeley, and University of Delaware, focused on the question of how children learn best. The network will investigate how to combine direct instruction with exploratory learning in ways that foster learning both immediately and for the long-term. Some previous research shows that direct instruction is best, but other research shows that exploratory learning is better. The debate over which approach is better reflects the absence of a theory of how to combine these methods to yield better learning outcomes than by either approach alone. The research components of this project will combine elements of these approaches to develop a theory of guided learning. The research will inform educational policy by developing novel methods and ways of thinking about education that foster immediate learning while also promoting the engagement and enjoyment that can drive future learning.

Debates in education and cognitive development have treated choice between learning methods, such as direct instruction and exploratory play, as an either-or question. The current research aims to integrate the strengths of these approaches and extend them into a coherent theory of guided learning. The approach capitalizes on recent advances in computational modeling that provide formal accounts of learning through direct instruction and exploration in probabilistic models of children's learning. Through this collaborative, an integrated theory that addresses what guided learning is, when to provide guidance, and why it works, will be developed. Behavioral experiments will investigate guided learning in children in the early school years. Experiments will contrast children's learning in guided settings with learning in direct instruction and exploration settings to experimentally test assumptions and predictions of the new computational model of guided learning. This research aims to develop novel educational methods founded on novel, mathematically rigorous models, toward the goal of informing policy choices in authentic educational settings.

The award is from the Science of Learning-Collaborative Networks (SL-CN) Program, with funding from the SBE Division of Behavioral and Cognitive Sciences (BCS), the SBE Office of Multidisciplinary Activities (SMA), and the EHR Core Research (ECR) Program.

Asking questions is a common way that families and teachers support children's learning, and different types of questions influence children's exploration and understanding. Yet we still know little about the role of questioning in STEM learning. This project investigates how and why certain types of questions may be better at fostering learning. In particular, in contrast to questions asked by a person seeking to learn something new (information-seeking questions), children may learn better from questions asked by someone who clearly knows the answer and intends to teach (pedagogical questions). This project bridges what we know about the development of children's thinking with research on how specific types of questions work to support young children's learning of science. To learn more about this, it examines the role of two types of questions across a range of scientific methods including experiments in laboratory settings, observation of children's conversations with parents and teachers in museums and classrooms, and interventions to test how types of questions work in the everyday activity of storybook reading. In addition to their importance for families, project findings will be valuable to preschool and elementary teachers and to educators who work in informal science settings such as science museums, zoos, and play centers. The project is supported by the EHR Core Research (ECR) program, which funds basic research that seeks to understand, build theory to explain, and suggest interventions (and innovations) to address persistent challenges in STEM interest, education, learning, and participation.

The project builds upon a long tradition of educational and developmental research on children's learning from questions in formal and informal STEM learning environments. Taking a new perspective, the project takes into account the relations among questions, the questioner, and learning goals. Specifically, it proposes that questions from knowledgeable others who intend to teach (i.e., pedagogical questions) may facilitate learning by eliciting children's inferences about the importance of the scientific subject being questioned, and by encouraging their further exploration and discovery learning. These assumptions will be tested across five empirical studies, each with its own specific aim. Aim 1 is to identify pedagogical questions and their effects on children's learning and exploration of scientific information in a controlled laboratory environment. Aim 2 seeks to characterize the use of pedagogical questions in math, physics, and biology classrooms using archival videos and transcripts from the ClassBank database, and explores the relation of pedagogical questions to student engagement. Aim 3 is to characterize the use of pedagogical questions in naturalistic parent-child conversations documented in the CHILDES (Child Language Data Exchange System) database, and explore the influence of socio-cultural factors on parents' tendencies to ask different types of questions. These factors include family socioeconomic status, presence of mothers and fathers, children's age, historical trend, and others. Aim 4 tests how natural variation in parents' use of pedagogical questions relates to children's learning and exploration in a science museum. Aim 5 tests an intervention that prompts parents to ask pedagogical questions during storybook sharing and examines how this intervention relates to children's learning of scientific knowledge. Across these five studies pedagogical questions are expected to have a facilitative role for children's learning and exploration, while socio-cultural factors are expected to influence how pedagogical questions are asked by parents and interpreted by children. Taken as a whole, this project has potential to deliver empirical evidence and a more comprehensive theory about when and why questions lead to learning. This research is important to society because it can lead to evidence-based recommendations about how to effectively use questions to enhance STEM teaching and children's learning.

This award to Rutgers University-Newark supports the acquisition and deployment of a High-Performance Computing (HPC) cluster (named PRICE) dedicated to research, teaching and societal outreach efforts. PRICE will have 60 general compute (CPU) nodes and one graphical processing unit (GPU) node as well as storage appropriate for the planned usage over the lifetime of the machine. The enabled research develops along three main directions: atomistic modeling, neuroscience, and data science. Some atomistic models enabled by PRICE will study the structure and dynamics of proteins to address questions related to diseases such as Alzheimer’s. New materials modeling and design is enabled both by PRICE and by the development of new quantum simulation methods. The enabled simulations will also regard new materials design by way of genetic algorithms. The enabled neuroscience research regards computational analysis of experimental data to understand brain function, connectivity, and human behavior. Enabled data science research includes the formulation of novel cooperative artificial intelligence (AI) algorithms that will improve the outcome of machine learning (ML) models of broad applicability. In addition to enabling new science, the project realizes several societal broader impacts including broadening HPC literacy of underrepresented minorities and training the future NJ workforce using HPC in the classroom and development of new undergraduate and graduate curricula.

PRICE will comprise 60 compute nodes (52 cores/node), 700 TB of redundant storage and one GPU node (4 GPUs/node) to be housed at Rutgers University-Newark. PRICE will enable several additional research projects carried out by the PI, co-PIs, and major users at Rutgers-Newark and NJIT. The GPU portion enables state-of-the-art molecular dynamics simulations that elucidate structure and dynamics of proteins for the understanding of diseases, such as Alzheimer’s. GPUs also enable the efficient and timely execution of cooperative AI algorithms aimed at improving predictivity. The CPU nodes will enable quantum simulations aimed at materials engineering through density-functional theory calculations. These simulations facilitate the development of quantum models based on density functional theory, its subsystem formulation (which parallelizes efficiently over PRICE’s low-latency network) as well as quantum mechanical frameworks based on the exact factorization of the Schrödinger equation for multicomponent systems. CPU and GPU nodes will enable data analysis associated with neuroscience experiments aimed at uncovering how the brain modulates behavior and vision-related tasks as well as the study of neuron connectivity to understand brain function. Data from these experiments is growing exponentially due to increased instrument data flow from new fMRI units and improved technology allowing recordings of local field potentials from hundreds on neurons. The project also enables new science as well as the realization of societal broader impacts, such as broadening high-performance computing literacy of underrepresented minorities, training the future NJ workforce and recruitment of new faculty.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.