Learning optimisation by high firing irregularity.

Area of Science:

• Computational neuroscience
• Artificial intelligence

Background:

• Leaky integrate-and-fire (LIF) neurons are fundamental units in computational neuroscience.
• Spiking neural networks (SNNs) offer a biologically plausible model for neural computation.
• Understanding the role of neural firing patterns, such as irregularity, is crucial for advancing SNNs.

Purpose of the Study:

• To investigate the functional significance of irregular spiking in LIF neural networks at high firing rates.
• To determine if irregular spiking enhances the learning capabilities of SNNs.
• To explore different mechanisms for inducing irregular spiking and their impact on learning.

Main Methods:

• Simulating a network of LIF neurons.
• Implementing two distinct methods to induce high firing rate irregularity: partial somatic reset and temporally correlated inputs.
• Training the SNN using reward-modulated spike-timing-dependent plasticity (R-STDP).
• Evaluating network performance on the XOR benchmark problem and a general-sum game.

Main Results:

• High firing irregularity, irrespective of its generation mechanism, significantly enhances the learning capability of the SNN.
• The network demonstrated improved performance in both the XOR problem and the general-sum game when employing irregular spiking.
• Reward-modulated STDP effectively leveraged irregular spiking for improved learning.

Conclusions:

• Irregular spiking at high rates plays a crucial functional role in optimizing the learning capabilities of SNNs.
• The findings suggest that the brain might utilize high firing irregularity as a mechanism for learning optimization.
• This research provides insights into the design principles for more efficient and capable artificial neural networks.