Learning to discriminate through long-term changes of dynamical synaptic transmission

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Synaptic plasticity, encompassing both short-term and long-term changes, is crucial for neural computation.
• The precise computational role of the interaction between short-term synaptic plasticity (STP) and long-term synaptic changes remains debated.
• Understanding this interplay is key to deciphering how neural circuits learn and adapt.

Purpose of the Study:

• To derive a learning rule for release probability and maximal synaptic conductance in a circuit model.
• To investigate the computational role of STP in modulating long-term synaptic changes for input discrimination.
• To explore how combined recurrent and feedforward connections facilitate learning from natural inputs.

Main Methods:

• Development of a circuit model incorporating both recurrent and feedforward connections.
• Derivation of a learning rule for synaptic release probability and maximal conductance.
• Utilizing computer simulations to analyze the model's performance with varying synaptic parameters.
• Employing a standard perceptron as a linear classifier to quantify performance improvements.

Main Results:

• Short-term synaptic plasticity nonlinearly expands the input space for linear classifiers.
• Random recurrent networks decorrelate this expanded input space.
• Simulations showed up to a 100% performance improvement in a perceptron due to STP-induced dimensional increase.
• Synaptic parameter distributions at capacity limits depend on excitation-inhibition balance.

Conclusions:

• STP acts as a nonlinear feature expander, enhancing the discriminative power of linear classifiers.
• The integration of STP with recurrent networks provides a mechanism for decorrelating expanded inputs.
• The model offers a novel interpretation for neuronal responses outside the classical receptive field as a result of stimulus space decorrelation.