![[Mitchell book cover]](http://www.cs.cmu.edu/~tom/mlbook.gif)
Machine Learning and Analyzing Human Brain Activity
Tom Mitchell, Monday, February 12, 4-5pm, Psychology 101.
Abstract: In recent years there has been a breakthrough in instruments for observing human brain activity, and even more recently machine learning methods have emerged as a valuable new approach to analyzing this data.
This talk will present our recent research exploring the patterns of human brain activity associated with the meanings of different words and pictures. For example, machine learning methods can be used to train classifiers to decode whether a person is reading a word about tools or buildings from the fMRI image of their brain activation. The same trained classifier can decode the semantic category of the stimulus whether it is an English word, a Portuguese word, or a line drawing of the object. We will describe efforts to use machine learning to study the neural representations of meaning in the human brain, including the challenge of dealing with this very hight dimensional, very sparse training data sets.
Affecting Behavior: Roles of Affect in Interactions among Situated
Embodied Agents
Matthias Scheutz, Monday, February 5, 4-5pm, Psychology 101.
Abstract: Affect seems to be deeply intertwined with many parts of the human cognitive architecture, and also that of many animals. Affective states like happiness, fear, anger, disappointment and many others routinely accompany cognitive processes in humans and can influence attention, problem solving, action selection, and, most importantly, social interactions.
In this talk, I will present an overview of our work on exploring and understanding possible roles of affect in embodied agents that are situated in cooperative and competitive multi-agent environments. Specifically, I will show that affective states can have both beneficial architecture-internal roles (e.g., in deciding what to do next) as well as useful social roles (e.g., for displaying agent-internal states that are indicative of the agent's behavioral dispositions). Throughout the presentation, I will draw examples from our agent-based simulation models in biologically motivated competitive multi-agent environments, where agents compete for limited resources in order to procreate and survive, and from our work on complex robotic architectures for human-robot interaction, where humans need to work with autonomous robots in teams to jointly achieve a task. The latter is of particular relevance to the emerging field of human-robot interaction (HRI) where models of human cognition and behavior can help in the design of robots that will be able to interact with humans in natural ways.
Interaction and Learning in Cognitive Robot Architectures
Jochen J. Steil, Monday, January 15, 4-5pm in
Psychology 101.
Abstract: The next generation of embodied cognitive robots will comprise advanced mechatronics for the body and its hands, multiple sensory channels, large data storage, and enormous processing power. The talk presents integrated architectures to create such systems in a human centered way, where the emphasis is on learning and interaction to make robots easily usable in every day life. It features multi-modal verbal, gestural, and visual communication for instructing the Bielefeld GRAVIS robot in anthropomorphic grasping. Further highlights are interactive online learning of visual objects and neural movement control of the humanoid Honda robot ASIMO. These applications demonstrate the power of learning in cognitive models and for providing a route towards more intelligent robots. More info on the Bielefeld Cognitive Robotics Program