Date: Saturday, July 5

Speaker: Terry Stewart

Title:  The Semantic Pointer Architecture

Date: Tuesday, July 15

Speaker: Anne Collins

Title: Computational models of structure learning in humans.

Abstract: Even without supervision or incentives, humans are able to acquire the complex rules that govern their environment; they tend to structure their policy into abstract hierarchical rules. In this talk, I investigate how and why individuals build such rules and propose a new reinforcement-learning model that accounts for human structure learning, and affords new predictions. I also present a neural network implementation of structure learning based on known connectivity between prefrontal cortex and basal-ganglia. Behavioral results confirm model predictions: subjects' structure learning is evidenced by their ability to generalize and transfer knowledge while learning new rules. EEG results also support our previous findings, as well as neural network predictions: markers of structured decision making predict individual differences in transfer and thus abstract structure learning. These findings confirm our model’s predictions, and show that the seemingly suboptimal strategy of building complex structure affords long-term advantages given the opportunity to generalize across contexts in various sorts of new situations.


Slides from the DNF14 topic area is found at the workgroup page.

Name: Gregor Schoener

Title: Neural Dynamics for Higher Cognition

Abstract: I will give a big picture perspective on what is a possible path to understanding higher embodied cognition based on neural principles. I'll start with Braitenberg vehicles and show that they involve a behavioral attractor dynamics from which behavior emerges in closed loop. Neural dynamics closes loops within the agent's nervous system. This enables detection and selection decisions, working memory and categorization, which emerge from instabilities of the neural dynamics. I will review the other ingredients needed on the path to higher cognition: neural oscillators for timing and coordination, conditions of satisfaction for sequences, steerable maps for coordinate and other transforms, associative learning and pattern learning for object learning and recognition, and neural operators for relational schemata. The "final frontier" on the path to higher cognition is the neural grounding of abstract rules that can be applied to new objects. I will contrast the population level neural dynamic approach to feedforward neural networks, to information processing and to the vector symbolic architecture.