Semi-orthogonal subspaces for value mediate a binding and generalization trade-off.

Area of Science:

• Neuroscience
• Cognitive Neuroscience
• Decision Neuroscience

Background:

• Making choices requires associating option values with selection actions, a process known as the binding problem.
• Neural mechanisms underlying how the brain solves this binding problem are not fully understood.

Purpose of the Study:

• To investigate whether neural population subspaces are used by the brain to solve the binding problem.
• To determine how neural populations encode value and location information during decision-making.

Main Methods:

• Monitored neural activity in five reward-sensitive brain regions of macaques performing a choice task.
• Analyzed neural population activity to identify how offer values and locations were represented in distinct subspaces.

Main Results:

• Neural populations in reward-sensitive regions encode left and right offer values in distinct subspaces, facilitating binding.
• Orthogonal encoding of sequential offer values in neural subspaces also supports binding.
• Behavioral errors correlated with spatial, but not temporal, neural misbinding.
• Errors increased for extreme (low/high) offer values compared to medium values.

Conclusions:

• The brain utilizes semi-orthogonal neural population subspaces to bind abstract value information to specific features like location.
• This subspace-based neural representation provides a mechanism for solving the feature binding problem in decision-making.