An LSTM-based adversarial variational autoencoder framework for self-supervised neural decoding of behavioral...

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Linking neural data and behavior is challenging due to difficulties in unsupervised behavioral analysis and subject-specific neural variations.
• Developing generalized neural decoding models requires extracting subject-invariant features.

Purpose of the Study:

• To present data-driven solutions for unsupervised behavioral data analysis and automatic label generation.
• To extract subject-invariant features for generalized neural decoding models.
• To improve cross-subject transfer learning in neural decoding.

Main Methods:

• An unsupervised autoencoder method for behavioral data transformation and cluster-friendly feature space generation.
• Iterative refinement of clusters using soft clustering assignment loss for improved feature representations.
• Adversarial learning with a long short-term memory-based adversarial variational autoencoder (LSTM-AVAE) to capture shared neural information across subjects.

Main Results:

• The proposed model achieved 89.7% accuracy in cross-subject neural decoding on mouse cortical recordings.
• The adversarial network effectively eliminated subject dependency in neural representations, improving transfer learning.
• LSTM-AVAE demonstrated effectiveness in capturing temporal dependencies within neural data.

Conclusions:

• The framework is effective for unsupervised clustering and label generation of behavioral data.
• High accuracy in cross-subject neural decoding indicates the model's potential for relating neural activity to behavior.
• The approach offers a robust solution for generalized neural decoding and understanding neural-behavior relationships.