blend-behavior-guided-neural

What problem does it solve?

Neural activity and behavior are often difficult to model together, especially when behavior data is unavailable at inference time. BLEND addresses this by treating behavior as privileged information during training and relying on neural activity alone for inference using privileged knowledge distillation.

Core Features & Use Cases

• Model-agnostic framework: works with diverse neural architectures.
• Privileged knowledge distillation: teacher uses behavior signals to guide the student from neural data.
• Flexible distillation strategies: supports multiple learning objectives and setups for neural decoding, neuron-type prediction, and brain-machine interface tasks.
• Real-world applicability: suitable for neuroscience experiments, brain-machine interfaces, and electrophysiology datasets where behavior data is limited.

Quick Start

Provide neural activity data (and optionally behavior signals) to train a BLEND model and perform inference using only neural activity.