Inferring relevance in a changing world.

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Psychology

Background:

• Reinforcement learning models typically focus on stimulus-action-reward associations.
• Real-world environments are multidimensional, with only a subset of features being relevant for rewards.
• Identifying relevant environmental dimensions is crucial for effective decision-making.

Purpose of the Study:

• To investigate the process of "representation learning" in humans.
• To understand how individuals learn to identify relevant environmental dimensions for reward.
• To compare human performance with normative models of representation learning.

Main Methods:

• Experimental task where one stimulus dimension was relevant for reward at a time.
• The relevant dimension changed unpredictably, introducing environmental uncertainty.
• Human performance was analyzed against strategies like selective attention and probabilistic inference.

Main Results:

• Human performance was better explained by selective attention and serial-hypothesis-testing.
• A normative strategy based on probabilistic inference did not accurately describe human behavior.
• The findings suggest humans use suboptimal strategies for learning relevant information.

Conclusions:

• Inferring relevance in complex scenarios may be computationally intractable for the brain.
• The brain employs approximation strategies, such as selective attention, for representation learning.
• These approximations are used even when optimal learning is computationally feasible.