When predict can also explain: Few-shot prediction to select better neural latents

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Latent variable models are crucial for understanding neural activity dynamics.
• Current prediction benchmarks like co-smoothing have limitations in assessing true dynamics.
• Lack of ground truth necessitates robust evaluation methods for these models.

Purpose of the Study:

• To reveal limitations of the co-smoothing prediction framework for latent variable models.
• To propose and validate a new metric, few-shot co-smoothing, for more accurate latent dynamics inference.
• To develop a novel validation measure for latent variable models in the absence of ground truth.

Main Methods:

• Utilized a student-teacher setup to demonstrate co-smoothing limitations.
• Introduced few-shot co-smoothing using regression on held-out neurons with fewer trials.
• Applied cross-decoding of latent variables from model pairs to identify minimal extraneous dynamics.

Main Results:

• Models with high co-smoothing can exhibit arbitrary extraneous dynamics.
• Few-shot co-smoothing effectively distinguishes between models with and without extraneous dynamics.
• A novel cross-decoding measure correlates with few-shot co-smoothing performance, validating the approach.

Conclusions:

• Co-smoothing alone is insufficient for evaluating latent variable model accuracy.
• Few-shot co-smoothing offers a more reliable metric for assessing latent dynamics.
• The proposed methods enhance the reliability of latent variable models for neural data analysis.