Engineering human intelligence by understanding how the brain learns
The key challenge for neuroscience and artificial intelligence
Machine learning algorithms have demonstrated an ability to succeed in various tasks, emerging as a general framework for problem solving. That being said, there is still a plenitude of brain’s advanced capabilities that the state-of-the-art algorithms fail to demonstrate.
Recent neuroscience studies have provided evidence to suggest that the brain has multiple separate modes of learning and inference about the world, each of which can guide behavior in unique ways. After decades of study, we begin to understand how the brain’s subsystems interact with each other to ultimately produce coherent behavior.
The next challenge for both artificial intelligence and decision neuroscience is thus to understand the neural circuitry that is flexible enough to perform a wide range of tasks. In this regard, it is becoming widely recognized that hierarchical control of learning systems may be both the way the human brain actually works and the optimal design for an artificial intelligence that operates under constraints on performance, time, and energy.
We aim to understand the nature of human intelligence on the deepest level, by interfacing neuroscience with artificial intelligence.
Recent studies have investigated neural mechanisms of different types of learning through a combination of various techniques measuring brain activity with computational learning models. However, little is known about how the brain determines which of these subsystems guides behavior at one moment in time. Our research interests are to develop a theory of how the brain, arguably at the higher level in the cognitive hierarchy of the prefrontal cortex, allocates control to multiple types of brain's subsystems, called meta-control. Specifically, we aim to investigate the neural and computational basis of the following functions:
- theory learning
- one-shot inference
- goal-driven learning
- metacognitive learning
We believe this understanding opens the possibility for making scientific and technological breakthroughs.
This theory will enable us to
(1-Neuroscience) understand the nature of human intelligence in the context of meta-control,
(2-Computational psychiatry) understand why and how a breakdown of the meta-control culminates in psychiatric disorders,
(3-Bioengineering) develop flexible and adaptive neuromorphic algorithms,
(4-Artificial Intelligence) design human-AI co-evolution engines with which humans and AI can learn from each other.