A compressed code for memory discrimination

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience

Background:

• Mnemonic discrimination, the ability to distinguish similar memories, is a key index of memory function.
• Traditionally, this is thought to involve neural code expansion for maximal distinctiveness.

Purpose of the Study:

• To test the alternative hypothesis that mnemonic discrimination is supported by lossy compression of visual input.
• To investigate whether discarding irrelevant details aids in distinguishing similar stimuli.

Main Methods:

• Trained neural networks to compress stimuli and measured representation "lossiness" using a mathematical framework.
• Empirically validated predictions using behavioral datasets and task fMRI data.
• Examined neural signatures of dimensionality reduction and information loss in visual and hippocampal regions.

Main Results:

• Greater "lossiness" in neural network representations predicted easier and better lure discrimination, supporting the compression hypothesis.
• Identified neural dimensionality reduction and information loss in higher-order visual stream (V4, IT) and hippocampal regions (dentate gyrus/CA3, CA1).
• These neural signatures correlated with mnemonic discrimination performance.

Conclusions:

• Mnemonic discrimination is supported by lossy compression, which efficiently encodes sensory information by discarding redundant details.
• This challenges the expansion hypothesis and highlights the role of efficient information processing in memory.
• Findings suggest that discarding information, rather than preserving all details, is crucial for distinguishing similar memories.