Reservoir Computing Properties of Neural Dynamics in Prefrontal Cortex

Area of Science:

• Computational neuroscience
• Cognitive science
• Neurobiology

Background:

• Primates adapt to novel situations using complex internal states beyond immediate sensory input.
• Cortical dynamics generate representations of these situations, with mixed selectivity neurons playing a key role.
• Randomly connected recurrent networks, like those in reservoir computing, naturally produce mixed selectivity.

Purpose of the Study:

• To explore the representational power and dynamics of reservoir networks in modeling cortical function.
• To investigate how mixed selectivity in these networks contributes to understanding complex cognitive tasks.
• To compare reservoir network dynamics with neural activity in primate brains.

Main Methods:

• Utilized reservoir computing, a framework modeling recurrent neural networks, to simulate cortical dynamics.
• Trained a reservoir network on a complex cognitive task designed for primates.
• Incorporated a feedback neuron to amplify context representations, mimicking learning effects.

Main Results:

• The reservoir model inherently exhibited dynamic mixed selectivity, crucial for representing behavioral context over time.
• Training a feedback neuron enhanced pre-coded context representations, improving model robustness.
• Reservoir activity dynamics closely matched neural activity observed in the monkey dorsal anterior cingulate cortex.

Conclusions:

• Reservoir computing provides a pertinent framework for modeling local cortical dynamics and their role in higher cognitive functions.
• Mixed selectivity in recurrent networks is fundamental for representing complex contextual information.
• Hybrid dynamical regimes combining reservoir processing and feedback neuron dynamics can solve complex cognitive tasks.