Turing universal neural networks do not require global clocks.

Area of Science:

• Computational neuroscience
• Theoretical computer science
• Artificial intelligence

Background:

• Recurrent neural networks (RNNs) were established as Turing universal in the 1990s.
• The rise of large-scale AI deployment highlights the critical need for energy efficiency.
• Previous research focused on synchronous neural network models.

Purpose of the Study:

• To extend the computational foundations of neural networks to asynchronous models.
• To investigate the computational power and efficiency of asynchronous neural networks.
• To address the perceived impracticality of asynchronous computing due to update variability.

Main Methods:

• Modeling asynchrony by updating a single, randomly selected neuron per step to eliminate global updates.
• Introducing specific design constraints to achieve Turing universal asynchronous architectures.
• Proving universality for both asynchronous fixed architectures with varying-precision neurons and variable architectures with fixed-precision neurons.

Main Results:

• Demonstrated that asynchronous neural networks can achieve Turing universality.
• Established universality for asynchronous fixed architectures with varying-precision neurons.
• Established universality for variable asynchronous architectures with fixed-precision neurons.

Conclusions:

• Asynchronous neural networks preserve full computational power, equivalent to synchronous models.
• Asynchronous architectures are amenable to efficient training methods.
• Asynchronous networks offer substantial potential for reducing energy consumption in AI systems.