Optimal neural inference of stimulus intensities

Area of Science:

• Computational neuroscience
• Machine learning
• Neural computation

Background:

• Natural data often exhibits correlations between stimulus class and intensity.
• Current machine learning models typically normalize inputs, ignoring these statistical properties.
• Biological mechanisms for neural circuits to handle stimulus dependencies in learning and inference are not well understood.

Purpose of the Study:

• To develop a probabilistic framework for modeling class-specific intensity variations in stimuli.
• To derive neural computation-inspired inference and learning rules.
• To investigate the role of intrinsic plasticity (IP) in sensory learning and representation of input statistics.

Main Methods:

• Utilized a probabilistic framework to model stimulus intensity variations.
• Derived approximate inference and online learning rules.
• Simulated a neural circuit with synaptic and intrinsic plasticity.

Main Results:

• Demonstrated that a neural circuit with specific plasticity can learn stimulus features and intensities concurrently.
• Provided a normative interpretation of intrinsic plasticity in sensory learning.
• Showed that neurons can represent complex input statistics through their excitabilities.
• Achieved improved statistical representations for visual and auditory data.

Conclusions:

• The proposed model offers a biologically plausible mechanism for learning from naturalistic stimuli.
• Intrinsic plasticity is crucial for sensory learning and representing non-trivial input statistics.
• The model has practical applications, such as estimating speech contrastive stress.