Gershman's 'Subjective Functions' lets AI define its own goals
A Harvard neuroscientist proposes that AI should invent its own objective functions like humans do.
In a new paper, Harvard scientist Samuel J. Gershman tackles a fundamental blind spot in AI: where do objective functions come from? Current systems optimize external task goals, but human intelligence constantly invents new objectives on the fly. Gershman proposes 'subjective functions'—higher-order objective functions that are endogenous to the agent, defined by the agent’s own features rather than external tasks.
The paper uses expected prediction error as a concrete example of a subjective function. This approach draws deep connections across psychology, neuroscience, and machine learning, suggesting that truly intelligent AI needs to be able to synthesize its own goals. The work could lead to systems that are more flexible, self-guided, and capable of open-ended learning—a step beyond today's reward-maximizing agents.
- Introduces 'subjective functions' as goal-generation mechanisms derived from agent features, not external tasks.
- Uses expected prediction error as a concrete example of how an agent can self-generate objective functions.
- Bridges AI with psychology and neuroscience to understand goal synthesis in human cognition.
Why It Matters
Could unlock AI that autonomously sets its own goals, moving beyond rigid task-specific optimization.