Attention maps reveal stimulus-dependent retinal population codes.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neuroscience

Background:

• Deep learning models require both predictive accuracy and interpretable mechanisms to map neural activity to stimuli.
• Understanding neural population codes is crucial for decoding sensory information.

Purpose of the Study:

• To investigate if transformer model attention mechanisms can provide biologically meaningful insights into neural population codes.
• To decode large-scale retinal ganglion cell recordings using a novel transformer architecture.

Main Methods:

• Employed the POYO framework, a scalable transformer architecture utilizing spike tokenization and latent modeling.
• Analyzed encoder and decoder attention patterns under uniform flash and structured moving ball stimuli.
• Utilized attention-guided ablations to validate the functional significance of identified neural subpopulations.

Main Results:

• The POYO framework reliably decoded both stimuli and rapidly adapted to new preparations, indicating transferable population codes.
• Encoder attention patterns adapted to stimulus complexity, showing synchronized/broad distribution for uniform stimuli and specialized allocation for structured stimuli.
• High-attention neurons exhibited distinct physiological signatures, and attention-guided ablations confirmed their causal role in decoding.

Conclusions:

• Generic attention mechanisms in transformer models can spontaneously recover biological coding strategies.
• These models can identify functionally distinct neural subpopulations without supervision.
• Transformer-based architectures are validated as a powerful tool for neuroscientific discovery.