Inferring population dynamics in macaque cortex.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Neural Engineering

Background:

• Multi-unit cortical recordings have increased, driving interest in neural population dynamics.
• Modeling neural population dynamics requires inferring unobserved neuron activity and predicting future states.
• A benchmark was established using macaque cortex neural recordings during motor tasks.

Purpose of the Study:

• To evaluate general-purpose recurrent neural networks (RNNs) against specialized models for neural population dynamics.
• To introduce a novel hybrid architecture (TERN) combining RNNs with self-attention for improved performance.
• To establish the state-of-the-art performance on the neural latents benchmark.

Main Methods:

• Utilized recurrent neural networks (RNNs) with masking for discriminative learning.
• Developed a novel hybrid architecture (TERN) augmenting RNNs with self-attention mechanisms.
• Evaluated models on four benchmark datasets of macaque cortical neural recordings.

Main Results:

• RNNs trained with masking outperformed all previously published models on all four benchmark datasets.
• The TERN hybrid architecture demonstrated further performance improvements.
• Pure transformer models did not achieve comparable performance levels.

Conclusions:

• The autoregressive bias inherent in RNNs is crucial for achieving peak performance in modeling neural population dynamics.
• The study establishes a new state-of-the-art on the neural latents benchmark.
• The authors propose augmenting the benchmark to include evaluations favoring generative models.