DeepInsight

DeepInsight decodes sensory and behavioral variables from raw, unsorted wide-band neural activity using deep learning to analyze neural representations without spike sorting.

Key Features:

• Deep learning-based decoding: Uses deep learning to map raw wide-band neural signals to sensory and behavioral variables.
• No spike sorting required: Operates directly on unsorted neural data, bypassing traditional spike sorting and other extensive preprocessing.
• Recording modality support: Applied to calcium imaging and electrophysiology recordings.
• Cross-condition generalization: Demonstrates robust generalization across stimuli, behaviors, brain regions, and recording techniques.
• Network interpretability: Enables analysis of the trained network to identify informative elements of the neural code.
• Empirical validation: Validated on rodent auditory cortex and hippocampus datasets, including identification of a head direction representation in putative CA1 interneurons.
• Objective decoding assessment: Provides an approach for measuring decoding performance directly from raw recordings.

Scientific Applications:

• Decoding sensory variables: Infers sensory-related signals from wide-band neural recordings.
• Decoding behavioral variables: Infers behavioral correlates, such as head direction, from unsorted neural data.
• Neural-code discovery: Identifies informative neural elements and novel representations, exemplified by putative CA1 interneuron encoding of head direction.
• Modality-comparative analysis: Enables comparative analysis of calcium imaging and electrophysiology data using a common decoding framework.
• Performance benchmarking: Assesses decoding performance objectively without relying on spike-sorted inputs.

Methodology:

Deep learning models are trained on raw, unsorted wide-band neural recordings and the trained networks are analyzed to identify informative elements of the neural code.