Fast sigmoidal networks via spiking neurons

Area of Science:

• Computational Neuroscience
• Artificial Neural Networks
• Biophysics

Background:

• Traditional neural network models often interpret neuron activity via firing rates.
• Biological neural systems exhibit complex temporal dynamics that are not fully captured by rate-based models.
• Understanding the computational principles of biological neurons is crucial for advancing both neuroscience and AI.

Purpose of the Study:

• To demonstrate that networks of realistic biological neuron models can simulate sigmoidal neural networks.
• To introduce a novel simulation approach based on temporal coding of neural signals.
• To explore the implications of this temporal coding approach for neural computation and hardware implementation.

Main Methods:

• Utilizing mathematical models of biological neurons capable of temporal coding (single spikes or synchronous firing).
• Developing a simulation framework that interprets neural information through spike timing rather than average firing rates.
• Analyzing the computational capacity of these spiking neuron networks as universal approximators.

Main Results:

• Networks of spiking neurons can effectively simulate arbitrary feedforward sigmoidal neural nets.
• This temporal coding approach leads to substantially faster simulations compared to rate-based methods.
• Spiking neuron networks demonstrate universal approximation capabilities for continuous functions using temporal coding.
• The proposed method is robust to noise and consistent with experimental findings on neural processing speeds.

Conclusions:

• Temporal coding offers a viable and efficient mechanism for computation in biological neural networks.
• This paradigm shift suggests new directions for developing learning rules in self-organizing neural systems.
• The findings pave the way for novel hardware implementations of neural networks using pulse stream Very Large-Scale Integration (VLSI).