Interpretable deep learning for deconvolutional analysis of neural signals.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuroscience

Background:

• Deep learning models neural activity but often act as "black boxes."
• Lack of interpretability hinders understanding the link between neural activity and network parameters.

Purpose of the Study:

• To develop an interpretable deep learning method for analyzing neural activity.
• To establish a direct interpretation of network weights in relation to neural activity using a generative model.

Main Methods:

• Algorithm unrolling to design sparse deconvolutional neural networks.
• Introducing deconvolutional unrolled neural learning (DUNL).
• Applying DUNL to deconvolve single-trial local signals across brain areas and recording modalities.

Main Results:

• Uncovered multiplexed salience and reward prediction error signals in dopamine neurons.
• Performed simultaneous event detection and characterization in somatosensory thalamus.
• Characterized heterogeneous neural responses in piriform cortex and striatum during naturalistic experiments.

Conclusions:

• DUNL provides a mechanistic understanding of neural activity.
• Advances in interpretable deep learning offer new insights into brain function.
• DUNL demonstrates versatility across different brain regions and recording types.