Machine Learning & Cognition
While more applications and technologies are integrating machine learning, fundamental issues remain with how machines make decisions, adapt decision-making based on experience, and function more like a human brain.
A. Lynn Abbott
A. (Louis) Beex
JoAnn M. Paul
For computers to understand what’s happening in a visual scene or image, they need to recognize how humans interact with surrounding objects. We are tackling the challenge of detecting human-object interactions, which will take us a step closer to a fine-grained visual understanding of human activities.
We are analyzing multi-channel electroencephalogram (EEG) data to discern the brain wave patterns associated with ADHD (attention deficit hyperactivity disorder). Using machine learning approaches on a small homogeneous dataset, we have been able to classify ADHD versus non-ADHD with very high accuracy. While this is a promising start, these approaches need to be applied to a much larger, non-homogeneous dataset in order to ascertain usability. Together with colleagues from the Department of Psychology, we are working toward establishing a large ADHD database so that our machine learning approaches can be refined and extended.
Dreaming may be the key to understanding how the brain enables massively higher levels of parallelism than are found in computers, as well as how creative processes occur in the brain. We are focusing on the brain’s ability to build and hypothesis-test potential strategies in nearly perfect parallelism, so that real-time (awake) usage models may be optimized. Computational dreaming (CD) is a meta-algorithm for cognition, and we have developed a maze-solving CD simulator that is about seven times superior to random model selection. Our work shows that CD tends to rediscover known maze-solving algorithms on its own, developing them from primitive moves to complete enumeration combined with a dream phase.
Biological systems usually perform well in the face of significant uncertainty, and ECE researchers are drawing on these systems, especially their control functions, to design artificial systems. We are investigating the theory of active inference, which seeks to explain a wide range of biological phenomena, as a means of controlling nonlinear systems in the presence of uncertainty and stochastic disturbances.
Active inference says that biological systems update a set of internal states that represent beliefs about the environment. The error between observations of the environment and predictions made from the internal states drives this process, updating beliefs about states and predictions, causing the system to act.