This cluster brings together AI, Psychology, and Neuroscience researchers dedicated to discovering the pillars of intelligence---the capabilities underlying intelligence as we know it. The goal is twofold: to understand and model natural forms of intelligence (human and animal) using tools from AI and to build AI grounded in the real world. Current popularized efforts toward Artificial General Intelligence (AGI), such as large language models, focus on scenarios detached from how humans and animals reason and behave in the real world. Text data may be easily accessible online to train large models, but language is merely the most recent addition to human intelligence. It relies on many layers of capabilities for which we still need sound models. We propose directly tackling the capabilities underlying natural intelligent systems, such as sensory-motor control, causality, social understanding, and innovation. Such capabilities rely on learning from multimodal, causal, and constantly changing data in complex real-world environments. These are the data types that humans and animals easily understand, but current AI efforts still cannot utilize.
To progress toward more accurate models of intelligence and toward AI agents that can operate in the real world, we must combine computational efforts with well-grounded theory and experimental evidence about natural forms of intelligence. For this, a genuinely interdisciplinary effort is a necessity. A core goal of this program is to jump-start long-term collaborations to contribute to the academic discourse in all three disciplines.
We focus on the following pillars of intelligence:
- Visual perception and reasoning
- Physical and causal reasoning
- Sensory-motor learning
- Active exploration, information-seeking, and curiosity
- Curriculum setting
- Social intelligence and learning from other agents
- Language production, understanding, and reasoning
- Creativity and innovation