Decoding and encoding (de)mixed population responses.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• The brain operates via large populations of interacting neurons.
• Advances in recording techniques provide unprecedented views into neural population dynamics.
• Analyzing large-scale neural data is complex due to mixed dependencies on experimental parameters.

Purpose of the Study:

• To review insights from dimensionality reduction methods for analyzing neural population activity.
• To demonstrate the stability and typicality of mappings between experimental parameters and neural population activity.
• To propose a framework for understanding population coding based on demixed latent signals.

Main Methods:

• Dimensionality reduction techniques to 'demix' neural activity dependencies.
• Analysis of large-scale neural datasets across various tasks, brain areas, and animals.
• Review of neural network architectures relevant to population coding.

Main Results:

• Mappings from experimental parameters to population activity are often typical, stable, and linearly identifiable.
• Dimensionality reduction effectively demixes complex dependencies in neural data.
• Latent signals (stimuli, decisions, etc.) are encoded non-linearly and decoded linearly.

Conclusions:

• Population coding represents demixed latent signals through neural activity.
• This coding scheme facilitates information propagation across cortical areas.
• Neural network architectures can bridge single-neuron biophysics and population activity.