The curse of dimensionality in motor cortex

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Understanding motor cortex function is crucial for neuroscience.
• Unsupervised dimensionality reduction, like principal component analysis (PCA), is common for analyzing neural data.
• Dimensionality is often assumed to be an intrinsic property of neural dynamics.

Purpose of the Study:

• To investigate how unsupervised dimensionality estimates behave with increasing neural recordings.
• To compare the effectiveness of supervised versus unsupervised methods in analyzing neural data at scale.
• To challenge existing assumptions about cortical dimensionality.

Main Methods:

• Utilized unsupervised dimensionality reduction techniques (e.g., PCA) on neural recordings from motor cortex.
• Analyzed neural data across four animals, scaling recordings up to 1000 electrodes.
• Trained decoders using both unsupervised subspaces and supervised methods.

Main Results:

• Unsupervised dimensionality estimates did not converge but scaled with the number of recorded neurons.
• Dimensionality increased with electrode count, showing non-saturating growth.
• Supervised decoders significantly improved with more electrodes, achieving high accuracy with <0.1% variance at 1000 electrodes.

Conclusions:

• Current unsupervised methods for estimating neural dimensionality are sensitive to recording scale, not just intrinsic neural properties.
• Supervised learning approaches effectively utilize large neural datasets for decoding movement.
• These findings necessitate a re-evaluation of computational methods in neuroscience as data volume increases.