DEMO focuses on origins and principles of organization for complex systems. If we can create chain reactions in complexity with a universal computational substrate, then certain hard problems should become easier to solve.

Exploiting the Complexity of Dynamical Systems

We have shown that recurrent neural networks can have extensive fractal state-spaces. This is exploitable in several ways for any general system where sensitivity to small changes requires amplified differences. We use such dynamical systems in our work on cognition, including as a model for highly generative reconstructive memory or mental imagery, increasing the capacity of our Recursive Auto-Associative Memory (RAAM) model for representing recursive structures in neural networks.

Co-Evolutionary Learning

Most learning takes place as optimization of a fixed environment or fitness function. This requires that the learner be "pre-adapted" to that environment in order to learn anything, or that the environment be "gradient engineered" for the particular learning mechanism. This inductive bias usually makes the researcher the primary cause of learning. We focus on dynamically-changing learning environments, often composed of competing learners, where the complexity of the task gradually and automatically increases without human intervention. We work in games, language learning, and computational optimization tasks.

Structure in Neural Networks

The fixed kinds of structures upon which neural network learning technology operates is a case where the network and the learning environment have been pre-adapted. Successful uses of neural networks have required humans to do the hard learning, and then use weight adjustment for fitting parameters. We look at problems where task-specific structure is necessary and can arise through modulated mutation of network structures themselves.

Evolutionary Machines

The long term goal is the co-evolution of machines and their brains, first in simulation, then, through advanced computer-aided manufacturing, into actual hardware. Our initial experiments simulate LEGO structures, evolve machine codes and interpreters, and groups of simulated robot agents.