A Variational Encoder Framework for Decoding Behavior Choices from Neural Data.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Decoding animal behavior from neural activity is crucial for understanding brain function.
• Existing methods often struggle with subject-specific variations in neural data.
• Developing robust, cross-subject decoding methods is a significant challenge.

Purpose of the Study:

• To propose a data-driven, two-step approach for extracting cross-subject neural representations.
• To decode subjects' behavioral choices from neural activity by overcoming individual differences.
• To leverage adversarial learning to create subject-independent feature representations.

Main Methods:

• Utilized an adversarial variational encoder model for representation learning.
• Employed a clustering model to label behavioral choices based on recorded characteristics.
• Applied an adversary network to remove subject individuality from neural representations.
• Recorded cortical activity from Thy1-GCaMP6s transgenic mice during a licking task.

Main Results:

• Successfully extracted discriminative, subject-independent feature representations from neural data.
• Achieved an average classification accuracy of 88.8% for decoding behavior choices across subjects.
• Demonstrated the effectiveness of the adversarial approach in generalizing across different subjects.

Conclusions:

• The proposed two-step adversarial variational encoder method effectively decodes behavior from neural activity.
• The approach successfully extracts generalized neural representations, mitigating subject-specific confounds.
• This method offers a promising direction for cross-subject neural decoding in neuroscience research.