In the Theater of Dreams - Adam Safron & Zahra Sheikhbahaee

With generous support from the California Institute for Machine Consciousness, our research program focuses on how models of purported "conscious functions" can be tested with hybrid AI/ML architectures. Here we will describe how augmenting a Dreamer world-modeling agent with a brain-inspired "global workspace" can effeciently compress competing sensory streams—such as vision, proprioception, and language—into a single coherent state for imagination and control. Ultimately, by forcing this compressed, shared state and grounding language instructions in observed environmental changes, our goal is to evaluate whether this consciousness-inspired architecture achieves better out-of-distribution generalization when faced with novel scenes and task combinations.