The brain's representations may be compatible with convolution-based memory models.

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Psychology

Background:

• Convolution models are effective for explaining memory associations.
• These models face challenges with auto-correlated, naturalistic item representations, questioning their neural plausibility.
• Current solutions involve using noise-like representations, which lack clear biological grounding.

Purpose of the Study:

• To investigate if a compromise exists between naturalistic and noise-like representations for convolution models.
• To explore the role of the brain's auto-correlation reduction mechanisms in enhancing representational compatibility.
• To support the neural plausibility of convolution-based association memory.

Main Methods:

• Applied Difference-of-Gaussians (DoG) transform to naturalistic images to reduce auto-correlations.
• Evaluated the performance of DoG-transformed images within convolution-based memory models.
• Compared the performance of DoG-transformed images against both naturalistic and noise-like representations.

Main Results:

• DoG-transformed images, while not as optimal as noise-like representations, were better preserved by the convolution model than raw naturalistic images.
• The brain's tendency to reduce input auto-correlations, exemplified by retinal processing (DoG), can yield representations compatible with convolution.
• This demonstrates a proof-of-principle for biologically plausible representations in association memory models.

Conclusions:

• The brain's auto-correlation reduction mechanisms may naturally produce representations suitable for convolution-based memory models.
• Difference-of-Gaussians (DoG) transformation offers a biologically plausible method to improve representational compatibility.
• This research strengthens the neural plausibility of convolution-based models for association memory.