A role for cortical interneurons as adversarial discriminators

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• The brain learns sensory representations through experience, but the underlying algorithms are not fully understood.
• Generative models and adversarial algorithms are leading computational theories of sensory learning.
• Linking these theories to specific brain cell types is crucial for understanding cellular mechanisms.

Purpose of the Study:

• To investigate the role of cortical interneurons in sensory learning.
• To determine if interneurons can function as discriminators in an adversarial learning algorithm.
• To explore the plasticity rules and computational properties of such a system.

Main Methods:

• Proposed a computational model where cortical interneurons act as discriminators in an adversarial learning framework.
• Characterized interneuron plasticity as switching between Hebbian (waking) and anti-Hebbian (dreaming) rules.
• Evaluated the model's performance in learning representations within recurrent neural networks.

Main Results:

• The proposed adversarial algorithm with interneurons as discriminators excels at learning representations in networks with recurrent connections.
• The algorithm's scalability is limited by network size.
• Oscillating activity between evoked states and generative samples can partially mitigate scalability issues.

Conclusions:

• Cortical interneurons could mediate sensory learning by acting as discriminators in an adversarial algorithm.
• This strategy involves state-dependent plasticity rules (Hebbian/anti-Hebbian).
• The model presents a plausible and testable hypothesis for biological sensory learning mechanisms.