Representation Biases: Variance Is Not Always a Good Proxy for Importance.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• Neuroscience commonly analyzes neural representations using methods like PCA and RSA.
• These methods often rely on the "linking assumption" that high-variance neural features are computationally critical.

Purpose of the Study:

• To challenge the linking assumption in neuroscience.
• To explore how representation biases from machine learning impact neural data analysis.
• To investigate the consequences of biased representations for understanding brain function.

Main Methods:

• Review of machine learning literature on representation biases.
• Theoretical analysis of how biases affect standard neuroscience analysis techniques.
• Conceptual case study using homomorphic encryption.

Main Results:

• Learned representations can be biased, overrepresenting simple features and underrepresenting complex ones.
• Standard analyses assuming high-variance features are key can lead to biased inferences about system simplicity and similarity.
• Critical computational mechanisms may reside in low-variance neural components.

Conclusions:

• The linking assumption in neuroscience is potentially flawed due to representation biases.
• Relying solely on high-variance signals can obscure important computational functions.
• A comprehensive understanding of neural systems requires analyzing all components, not just the most prominent.