Learning what matters: Synaptic plasticity with invariance to second-order input correlations.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Cortical neurons form sparse representations reflecting environmental statistics.
• Synaptic plasticity must distinguish true features from spurious input correlations for efficient coding.
• Existing models struggle with omnipresent second-order correlations in cortical networks.

Purpose of the Study:

• To develop a theory for sparse coding and synaptic plasticity invariant to second-order input correlations.
• To explain the role of Hebbian long-term depression (LTD) in refining neural representations.
• To enhance biologically plausible learning models for improved decoding and sparse coding.

Main Methods:

• Developed a novel theoretical framework for sparse coding and synaptic plasticity.
• Analyzed the functional form of excitatory plasticity, including Hebbian LTD.
• Utilized a spiking neural network model with triplet spike-timing-dependent plasticity (STDP).

Main Results:

• The theory demonstrates how LTD cancels sensitivity to second-order correlations, aligning receptive fields with higher-order statistics.
• Invariance to second-order correlations improves decoding from noisy and spatially correlated inputs.
• Individual neurons in the STDP model learned localized, oriented receptive fields without preprocessing or lateral inhibition.

Conclusions:

• The developed theory advances understanding of local unsupervised learning in cortical circuits.
• Hebbian LTD plays a crucial role in enabling neurons to learn relevant features from complex sensory data.
• The findings offer new interpretations for models like Bienenstock-Cooper-Munro and triplet STDP, highlighting LTD's function in pyramidal neurons.