Tool use is an advanced behavioral capability, only observed in a small subset of species, crucial to their ecological success, shaped by their evolutionary histories, and dependent upon an interconnected set of more basic cognitive components. A new project from a team led by Joshua Tenenbaum and Kevin Smith at Massachusetts Institute of Technology (MIT) and the Center for Brains Minds and Machines will test how those subcomponents are structured and combined across apes, parrots, and humans (adults and children). They will test the respective contributions of bottom-up and top-down processing across species in this evolutionarily important context.

This project is synergistic with the Framework designed by Foster and colleagues. Foster et al.’s
Framework proposes “ a set of formal models, which instantiate mechanisms that range from top down to bottom-up” and hypothesizes that “bottom-up mechanisms should appear when systems have less experience with a problem, while top-down mechanisms should dominate when a problem is familiar.”

This research will test this hypothesis and contribute to the Framework in three ways:

• Provide an evolutionarily important context (tool use) within which to test the Framework’s predictions about the relationship(s) between components and control.
• Provide a concrete model of problem solving with tool use (Sample, Simulate, Update or “SSUP model”) that instantiates both top-down and bottom-up processing.
• The team’s Virtual Tools platform will allow the researchers to explicitly test the Framework’s prediction that novel environments elicit more bottom-up processes, which switch to top-down processes as experience is gained. The Virtual Tools platform requires placing a “tool” to enact physical events in order to accomplish a goal. The team will create different game-like rounds organized into levels. When players encounter the first round in a level, the problem will be novel, requiring a new type of tool or application. Tool strategies will effectively transfer across rounds within levels, allowing the team to measure learning.

The project will proceed in three phases.

• In Phase 1, the team will test whether apes transfer real-world physical knowledge into the Virtual Tools environment, and thus whether they can compare their performance with human subjects directly.
• In Phase 2, they will compare the behavior of apes, parrots, and humans in novel virtual tool-use scenarios, examining similarities and differences across the subcomponents proposed by the SSUP model, or how those components are used.
• In Phase 3, the team will study how novel tool-use strategies are learned across species, thus allowing us to investigate how the contribution of bottom-up and top-down processes changes with experience.