Direct Discriminative Decoder Models for Analysis of High-Dimensional Dynamical Neural Data.

Area of Science:

• Integrative neuroscience
• Computational neuroscience
• Machine learning for neuroscience

Background:

• High-dimensional neural recordings require scalable and computationally tractable data analysis methods.
• Existing latent process models struggle to scale and often ignore crucial local neural activity history.
• Accurate inference of cognitive processes depends on analyzing neural data effectively.

Purpose of the Study:

• Propose a novel, scalable latent process model for direct estimation of cognitive process dynamics.
• Introduce the direct discriminative decoder (DDD) model, which bypasses the need for individual neuron receptive field models.
• Develop a flexible framework applicable to various neural data modalities.

Main Methods:

• The direct discriminative decoder (DDD) model incorporates a discriminative process and a state transition model.
• The discriminative process models the conditional distribution of the state using neural activity and its local history.
• The framework allows flexible choices for the discriminative process, including deep neural networks and Gaussian processes.

Main Results:

• The DDD model offers a computationally tractable solution for high-dimensional neural data analysis.
• The model effectively incorporates neural activity history across multiple timescales.
• An extension, the discriminative-generative decoder (DGD), integrates physiological correlates like behavior with neural data for enhanced cognitive process estimation.

Conclusions:

• The DDD and DGD methods provide significant computational and modeling advantages for analyzing complex neural recordings.
• These scalable models are powerful tools for advancing integrative neuroscience research.
• The proposed framework enhances the ability to infer cognitive processes from diverse neural data types.