A Mixture of Generative Models Strategy Helps Humans Generalize across Tasks

E-print/Working paper (2021)

What role do generative models play in generalization of learning in humans? Our novel multi-task prediction paradigm—where participants complete four sequence learning tasks, each being a different ... [more ▼]

What role do generative models play in generalization of learning in humans? Our novel multi-task prediction paradigm—where participants complete four sequence learning tasks, each being a different instance of a common generative family—allows the separate study of within-task learning (i.e., finding the solution to each of the tasks), and across-task learning (i.e., learning a task differently because of past experiences). The very first responses participants make in each task are not yet affected by within-task learning and thus reflect their priors. Our results show that these priors change across successive tasks, increasingly resembling the underlying generative family. We conceptualize multi-task learning as arising from a mixture-of-generative-models learning strategy, whereby participants simultaneously entertain multiple candidate models which compete against each other to explain the experienced sequences. This framework predicts specific error patterns, as well as a gating mechanism for learning, both of which are observed in the data. [less ▲]

A Formal Framework for Structured N-Back Stimuli Sequences

Scientific Conference (2019, September 15)

A generalizable performance evaluation model of driving games via risk-weighted trajectories

in Proceedings of The 12th International Conference on Educational Data Mining (EDM 2019) (2019)

Efficient learning experiences require content to dynamically match a learner's skill; this assumes a fast and accurate assessment of the learner's skill and the ability to update content accordingly ... [more ▼]

Efficient learning experiences require content to dynamically match a learner's skill; this assumes a fast and accurate assessment of the learner's skill and the ability to update content accordingly. Effective personalized learning therefore involves deriving a performance-predictive mapping between behavioral and environmental factors. Once learned, this relationship can be used to generate new content and to update skill estimates based on the learner's interactions in an adaptive system. To provide proof of concept: (1) We develop a fast-paced driving video game where the player skillfully navigates a cluttered environment comprising obstacles and collectibles. Game content is generated procedurally and player behavior is recorded in the game-this provides an ideal test-bed for a method aiming to learn such a performance-predictive mapping. (2) Using blurred occupancy maps of the game's segments, we generate risk-weighted trajectory profiles for each user and segment of the game. Here, we show that these profiles can be used in a regression model to predict in-game performance both within and between game segments. Additionally, these profiles themselves reveal a trade-off between in-game rewards and risks. Successful identification of predictive environmental units within the game provides insight into the mapping between environmental features and performance, while facilitating the process of procedurally generating new, appropriate content in our adaptive system. We show that rapidly assessed measures of risk are highly predictive of both driving performance and reward rate, providing proof-of-concept evidence for the feasibility of a personalized adaptive learning system for this game. [less ▲]

A Multi-Objective Optimization Algorithm to Generate Unbiased Stimuli Sequences for Cognitive Tasks

Poster (2019)

Cognitive scientists want to ensure that particular cognitive tasks target particular cognitive functions that can be mapped to stable neural markers. Numerous cognitive tasks, like the n-back, involve ... [more ▼]

Cognitive scientists want to ensure that particular cognitive tasks target particular cognitive functions that can be mapped to stable neural markers. Numerous cognitive tasks, like the n-back, involve generating sequence of trials which satisfy certain statistical properties.The common approach to generate these sequences however lacks a theoretical framework and induces unintentional structure in the sequences which affects both behavioral performance and might bias the people’s cognitive strategies when completing a task. For example, people might exploit local properties in a random sequence in their decision making process. We argue that optimized experimental design requires cognitive tasks to be served by stimulus sequence generators that satisfy multiple constraints, both at the global and at the local structures of the sequence and that these sequence properties need to be systematically incorporated in the behavioral data analysis pipeline. We then develop a framework to reformulate the sequence generation process as a compositional soft constraint satisfaction problem and offer a multi-objective, genetic-algorithm-based method to generate controlled sequences under behavioral and neural constraints. This approach provides a systematic and coherent framework to handle stimulus sequences which in turn will impact the insights that can be gained from the behavioral and neural data collected on people performing cognitive tasks using those sequences. [less ▲]