Inverted Encoding Models Reconstruct an Arbitrary Model Response, Not the Stimulus.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Sensory Perception

Background:

• Understanding neural population activity is crucial for deciphering sensory representations.
• Inverted encoding models (IEMs) have been used to infer population-level stimulus representations from human brain activity.
• However, the channel response functions derived from IEMs may not accurately reflect true population tuning.

Purpose of the Study:

• To investigate the validity of channel response functions derived from IEMs for assaying population tuning.
• To determine if IEMs can uniquely describe population-level stimulus representations.
• To propose a salvageable method for assessing population selectivity using IEMs.

Main Methods:

• Mathematical derivation to analyze the uniqueness of channel response functions.
• Simulations using bimodal and random channel basis functions to test model robustness.
• Modifying the IEM approach to reconstruct the stimulus instead of the model's hypothetical responses.

Main Results:

• Channel response functions from IEMs are arbitrary and not unique, being determined only up to an invertible linear transform.
• Simulations show that arbitrary basis functions can perfectly explain population responses without true neural tuning.
• Reconstructing the stimulus, even with arbitrary basis functions, successfully recovers a unimodal function indicating population selectivity.

Conclusions:

• The standard IEM approach does not uniquely assay population tuning and can yield arbitrary results.
• A modified IEM approach, focusing on stimulus reconstruction, provides a valid measure of population selectivity.
• This highlights a general challenge in interpreting complex analyses of neural population data.