System-Level Modeling of Intelligent Behavior
While many specific aspects of neural function have been modelled, it has sometimes been difficult to interconnect them. Jonathan Cohen’s team has constructed a new tool for modelling that will make it easier to do so, and thereby advance our understanding of the mechanisms underlying human intelligence and allow researchers to draw direct comparisons with artificial intelligence.
The project proceeds from two assumptions:
1) understanding human intelligence will be informed by understanding how this arises in the human brain, still the benchmark device for “general intelligence;”
2) such progress requires extensive use of computational modeling.
The latter is necessary for not only addressing the complexity of brain function but also expressing it in a form that enables direct comparisons with forms of artificial intelligence. Such strategies have already led to progress at the interface between neural network modeling in neuroscience and deep learning in computer science.
This project addresses two challenges presented by these goals:
1) the need to integrate knowledge gained about brain subsystems—responsible for domain-specific forms of intelligence—into system-level models of their interaction that can explain domain-general forms of intelligence; and
2) the need for software suitable not only for creating such system-level models, but also documenting, disseminating, validating, comparing, and integrating such models, including ones from machine learning that use AI.
The team will use PsyNeuLink, a tool recently created for this purpose. With PsyNeuLink, they will implement existing theories of well-characterized subsystems.
They will also generate computer simulations that illustrate how interactions among subsystems can give rise to three constituents of intelligent behavior at the system level:
1) explore-exploit decisions
2) planning, and
3) the strategic allocation of cognitive control.
To make both the tool and models as useful as possible, the team will install a graphical user interface, and add the new models to its open source repository.